We talk to companies fairly often after a failed AI engagement, not before one. By the time we hear about it, the pattern is usually the same: an impressive demo, a signed contract, and six months later, nothing in production. This is not rare. A recent RAND analysis found that more than 80 percent of AI projects fail to deliver, and separate research from S&P Global put project abandonment at around 42 percent. The technology is rarely the reason. The vendor selection is.
Choosing an AI consulting company is closer to hiring a strategic partner than buying software, and treating it like a software purchase is where most of these engagements go wrong. Here is what we think actually matters when you are comparing options.
Why this decision matters more than it used to
The pace of change is part of why vendor selection has gotten harder. This TEDx talk from physicist and technology researcher Alexander Wissner-Gross puts the scale of what is coming into perspective, and it is a useful piece of context before you sit down with any vendor.
Start with a one page scope brief, before talking to anyone
The quality of a vendor’s response tells you more than any sales demo, but only if you give them something specific to respond to. Before the first call, we recommend answering five questions internally: what problem are we solving, which business functions does it touch, what does success look like in 90 days, is our data actually ready, and how much internal bandwidth do we have to support the engagement.
A vague brief gets a vague, impressive sounding proposal from almost anyone. A specific brief filters out the vendors who are not a real fit before you waste a call on them.
The four types of firms, and why the distinction matters
- Big four and MBB-style firms (Deloitte, KPMG, McKinsey’s QuantumBlack and similar). Strong for board level credibility and cross-functional scale, engagements typically run 8 to 12 months and start around $150,000, often climbing well past $500,000. Junior consultants tend to handle day to day delivery.
- Specialist AI consultancies. Narrower focus, usually faster to engage, often 40 to 60 percent lower day rates than the large firms, with implementation as the core offer rather than strategy decks.
- Systems integrators. Strong on technical integration with existing infrastructure, sometimes weaker on the strategic “should we even do this” layer.
- Boutique or solo practitioners. Can move fast and offer real hands-on expertise, but capacity and continuity risk is higher if the engagement scales.
None of these categories is automatically the right answer. The mismatch between category and need is one of the most common reasons engagements stall.
What actually separates a good AI consulting company from a risky one
- Production track record, not just prototypes. Ask for examples of systems running in production today, not proof of concepts that never shipped.
- Discovery quality. A firm that runs a structured discovery phase, one that produces a risk register and a validated data assessment rather than a generic template, is far more likely to deliver. According to Gartner’s April 2026 survey, only 28 percent of AI use cases fully meet ROI expectations, and weak discovery is consistently one of the reasons the rest fall short.
- Technology-agnostic recommendations. Strong teams recommend platforms based on your requirements, not their existing commercial partnerships. Ask directly how they evaluate competing models before recommending one.
- Clear IP ownership. Get in writing who owns the models, the code, and the underlying data pipeline once the engagement ends.
- Knowledge transfer built into the plan. The engagement should leave your internal team able to run, monitor, and extend the system without the consultant in the room.
- References from companies your size and sector. General case studies from unrelated industries are weak evidence, push for examples that match your scale and problem.
Red flags worth ending the conversation over
- They lead with a specific tool or platform before understanding your problem.
- They claim to handle every department and every use case equally well.
- They cannot name a specific project, with a named business outcome, that they personally delivered.
- Pricing is vague or bundled in a way that makes it hard to compare against other proposals.
- No references from businesses similar to yours in size, sector, or challenge.
Start with a pilot, not a transformation
Before committing to a large scale engagement, validate the relationship with a smaller, time bound pilot that has a fixed budget and agreed success criteria. Check references directly, ask past clients what went well, what did not, and whether they would hire the firm again. Evaluate the relationship during the pilot too, not just the output, communication quality and transparency matter as much as technical performance once the contract is signed.
Frequently asked questions
How much does AI consulting typically cost?
Initial diagnostics for mid-market companies typically run $10,000 to $30,000 over 2 to 6 weeks. Full engagements vary widely by firm type, from tens of thousands with a specialist boutique to well over $500,000 with a large firm on an enterprise-wide rollout.
Should I choose a big name firm or a boutique consultancy?
Brand name does not predict implementation quality. A smaller firm with a strong production track record in your specific industry often outperforms a large firm that is learning your sector from scratch, at a fraction of the cost.
What is the biggest reason AI consulting engagements fail?
Weak discovery. Firms that skip straight to building, without a structured assessment of your data, workflows, and actual bottlenecks, tend to deliver systems that do not fit how the business actually operates.
How long should a first engagement be?
We recommend starting with a defined pilot, typically 4 to 8 weeks, with a fixed budget and clear success criteria, before committing to a larger scope.
What questions should I ask every vendor?
Ask them to describe one specific project they personally built or led, with a named business outcome. Ask who owns the models and data once the engagement ends. Ask how they evaluate competing AI models before recommending one.
Conclusion
The vendor determines whether your team finds out an approach will not work before signing the contract, or after spending the budget. A structured evaluation, a scoped pilot, and real reference checks catch most of the expensive mistakes before they happen.
Saiba mais em:
Gen AI Studio