I’ve lost count of how many AI vendor proposals I’ve reviewed over the past decade. Started in the Big 4, now I do it independently for enterprise clients across Australia and Southeast Asia. Conservative estimate? North of 400.

And I can tell you this: the proposals that end up failing in practice almost always show the same warning signs on paper. Here are the five red flags I look for now — and the ones that should make you pause before signing anything.

1. Vague ROI Promises With No Baseline Measurement

This one shows up constantly. The proposal says something like: “Our solution will deliver 30-40% efficiency improvements across your operations.”

Sounds great. But 30-40% of what, exactly? Against which baseline? Measured how? Over what timeframe?

I reviewed a proposal last year for a logistics company in Sydney. The vendor promised $2.4 million in annual savings. When I asked how they calculated that number, the answer was essentially “industry benchmarks.” They hadn’t looked at the client’s actual cost structure, hadn’t mapped their current processes, and hadn’t even asked what systems were already in place.

According to Gartner, through 2025 at least 30% of AI projects were abandoned after the proof-of-concept stage — and unrealistic ROI expectations were a leading cause. That number hasn’t improved much since.

If a vendor can’t explain exactly how they’ll measure success using your data and your current performance numbers, that’s your first red flag. Legitimate proposals include a discovery phase specifically designed to establish baselines before making promises.

2. “Our Proprietary AI” Without Technical Specifics

I get nervous whenever I see the phrase “our proprietary AI platform” without any further detail. What model architecture? What training data? What’s the tech stack underneath?

Sometimes this is just marketing fluff wrapping a perfectly fine product. But more often, it’s hiding something. Maybe they’re reselling an open-source model with a nice front-end. Maybe they’re using API calls to a third-party LLM and marking it up 400%. I’ve seen both.

One proposal I reviewed described their “proprietary machine learning engine” in glowing terms across three pages. When we got to technical due diligence, it turned out they were running a lightly customised version of XGBoost — a free, open-source library — inside a Docker container. The $850,000 price tag suddenly looked very different.

Ask for architecture diagrams. Ask what happens to your data. Ask where the model runs. If they dodge those questions, walk away.

3. No Discussion of Data Requirements

This might be the most telling red flag of all. A vendor who doesn’t ask about your data situation early and often is a vendor who’s going to cause you problems.

Good AI needs good data. That’s not a cliché — it’s the single biggest determinant of project success. And yet I regularly see proposals that dedicate 15 pages to features and capabilities and half a page to data integration requirements.

Here’s what I want to see in a serious proposal: What data formats do they need? What volume? What quality thresholds? How do they handle missing or messy data? What’s the data preparation timeline, and who’s responsible for it?

A manufacturing client I worked with signed a $1.6 million contract for a predictive maintenance system. Six months in, the project was stalled because the vendor hadn’t accounted for the fact that the client’s sensor data was stored in three different legacy systems with incompatible formats. That should have been identified in week one of scoping, not month six of delivery.

4. No Change Management Plan

Technology doesn’t fail in a vacuum. It fails because people don’t use it, don’t trust it, or actively work around it.

Any AI deployment that touches how people do their jobs needs a change management strategy. I’m talking about training plans, communication timelines, feedback mechanisms, and executive sponsorship structures. The whole thing.

I’ve seen $3 million AI platforms go completely unused because nobody thought to involve the people who’d actually need to interact with them every day. The vendor built exactly what was specified, delivered on time, and then watched as adoption flatlined at 12%.

When I see a proposal that’s all technology and zero people strategy, I flag it immediately. The best vendors I work with dedicate 15-20% of the total project budget to change management. It’s not optional — it’s foundational.

5. Firm Timeline Guarantees Without Caveats

“Full deployment in 12 weeks.” No asterisks. No dependencies listed. No mention of what could go wrong.

That’s not confidence — that’s recklessness. Or worse, it’s a vendor who plans to declare victory at the 12-week mark regardless of whether the thing actually works.

Honest timelines come with conditions. They say things like: “12 weeks assuming data access is available by week 2” or “16-20 weeks depending on integration complexity identified during discovery.” That’s what real project planning looks like.

I worked with a financial services firm that chose a vendor partly because they promised the fastest timeline. The project ran 14 months over schedule. The competing vendor, who’d quoted 8 months with clear dependencies, would have been done in 9.

The Bottom Line

None of these red flags are automatic disqualifiers on their own. Sometimes a vendor has a great product but a bad proposal writer. It happens.

But if you’re seeing three or more of these in a single proposal? That’s a pattern. And patterns in proposals tend to become patterns in delivery.

My advice: slow down the procurement process. Build in a proper technical evaluation phase. And don’t let impressive slide decks substitute for answering hard questions. The money you spend on due diligence is a rounding error compared to the cost of a failed AI project.