AI Consulting Firms: How to Choose When AI Labs Advise Too

AI Consulting Firms: Choosing an Adviser in a Crowded Market

AI consulting firms used to mean the big professional services houses and a scattering of specialist boutiques. That map is changing quickly. In late October, City AM reported that Anthropic is preparing to launch a consultancy arm in London with private equity firm Hg, with a model that appears to involve embedding AI engineers inside client organisations. The same week, Accenture formed an Advanced Intelligence Group drawing on more than two thousand specialists to tackle complex client AI work. And fresh research continues to suggest that most companies have yet to see AI improve their bottom line. For buyers, more choice is welcome. It also sharpens a question that matters more now than it did a year ago: when your adviser also sells the model, whose interests does the recommendation serve?

Types of AI Advisers

Each type of adviser brings different strengths and different conflicts.

Why the AI Consulting Market Just Changed

AI labs have strong reasons to move into advisory work. Deploying AI well inside a real organisation is hard, and many clients struggle to turn access to a model into business results. Advisory and engineering services help clients succeed, which drives usage and revenue, and it helps offset the enormous costs of running and training models. There is a historical parallel: enterprise software vendors built large consulting practices in the 1990s for similar reasons.

Investors so far seem relaxed about the threat to established firms, judging by the muted market reaction. But for clients, the arrival of model makers as advisers changes the buying conversation. A lab's engineers will know their models better than anyone. They will also, understandably, design solutions around them.

The Neutrality Question

Neutrality has always been an issue in consulting. Large firms maintain alliances with software vendors, and those relationships can shape recommendations. The new element is that the adviser and the model provider may now be the same company. That does not make the advice bad. It does make independence something a client should test rather than assume.

When AI Consulting Firms Also Sell the Model

The practical risks are specific. A design may assume one provider's model when another would be cheaper or better for part of the workload. Architecture choices may make switching providers harder later. Evaluation may be run only against the adviser's own models. None of this requires bad faith. It follows naturally from expertise and incentives. Clients can manage it by asking for model-agnostic architecture, independent evaluation against alternatives, and contract terms that preserve portability.

The Embedded Engineer Model

Much of the new competition rests on a delivery style borrowed from data and software companies: engineers who sit inside the client's teams, build alongside them, and iterate on real problems rather than producing recommendations from outside. It can work very well. Problems surface faster, solutions fit actual workflows, and less is lost between a strategy document and a working system. It also changes what clients need to manage. Embedded engineers make many small design decisions every week, and each one can deepen dependence on a particular model, tool, or way of working. Clients using this model should assign their own technical lead to review architecture choices regularly, and keep a running record of decisions so the system remains understandable after the engineers leave.

How to Choose: Seven Questions

1. What have you delivered that is in production? Ask for live systems and measured outcomes, not pilots and demos.
2. Who will actually do the work? Meet the delivery team, not just the people in the pitch.
3. Which model providers do you work with, and on what terms? Ask about alliances, revenue sharing, and incentives.
4. How will you evaluate alternatives? Insist on testing at least two or three models on your actual tasks.
5. How will you transfer knowledge? Your team should be able to run and improve the system after the engagement ends.
6. Who owns what you build? Code, prompts, evaluation data, and configurations should belong to you.
7. How do we switch providers later? Ask how the architecture supports changing models without a rebuild.

Structuring the Engagement

The contract shapes behaviour as much as the proposal does. In our view these elements make AI engagements more likely to deliver:

• Outcome metrics agreed up front. Define business measures such as cost per case, cycle time, or error rate, not just deliverables.
• Staged commitments. Release budget in phases tied to evidence rather than committing to a full programme at once.
• Independent evaluation. Require model comparisons and accuracy testing that your team can rerun.
• Embedded client staff. Pair your people with the adviser's engineers so capability stays in the organisation.
• Clear ownership of assets. Prompts, pipelines, evaluation sets, and documentation transfer at the end.
• Portability clauses. Avoid exclusive dependencies on one model provider unless there is a clear reason.

Pros and Cons of Lab-Run Consultancies

• Pro: deep technical expertise. Engineers who build models understand their strengths and limits intimately.
• Pro: early access. Clients may benefit from new capabilities before they reach general release.
• Pro: hands-on delivery. Embedded engineers can move faster than layered consulting teams.
• Con: model bias. Solutions naturally centre on the lab's own models.
• Con: limited industry depth. Domain knowledge and change management may be thinner than at established firms.
• Con: concentration risk. Advice, platform, and model from one source deepens dependency.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how choices between AI consulting firms play out in practice rather than presented as verified case studies.

A bank hires a model provider's advisory team to build a document review system. The system works well. A year later, a competing model handles part of the workload at a fraction of the cost, but the architecture makes switching expensive. The bank's next contract requires a model-agnostic design from day one.

A manufacturer runs a short competition between a large consultancy, a boutique, and a lab-backed team, giving each the same task with the same data. The boutique's solution is less polished but cheaper to run and easier for internal staff to maintain. The manufacturer chooses it, and keeps the evaluation set for future comparisons.

A retailer insists that its own analysts sit alongside the adviser's engineers throughout an engagement. When the contract ends, the internal team takes over and continues improving the system, avoiding a costly follow-on contract.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

Measuring Whether It Worked

Recent research from large firms suggests most companies have not yet seen AI translate into better financial results, even after significant investment. That is a reminder to measure AI engagements on outcomes rather than activity. Before work starts, record the baseline: how long tasks take, what they cost, how many errors occur. After deployment, measure the same things. If the numbers do not move, a polished deliverable is not a success.

Quality of the work itself also matters. Consulting firms have faced embarrassing incidents where AI-generated material slipped into client deliverables unchecked, a pattern we examined in AI in consulting. And the long-term risk of depending on one provider is a board-level question, which we covered in AI vendor lock-in.

Why Talkory Wins

The simplest test of an adviser's neutrality is to compare models yourself. Talkory runs the same question across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in one pass, giving clients a quick, independent view of how different models handle their actual tasks. If one provider is clearly best, the comparison will show it. If several perform similarly, that is a strong reason to insist on a model-agnostic design and keep switching options open. Either way, the client goes into the conversation with evidence rather than relying on the adviser's preference.

Final Verdict

AI consulting firms now include large consultancies, specialist boutiques, software vendors, and the AI labs themselves. That variety is good for clients, provided they buy carefully. Choose on delivered production outcomes, meet the real delivery team, ask about alliances and incentives, insist on independent model evaluation, secure knowledge transfer and ownership of what is built, and preserve the ability to switch providers. When the adviser also sells the model, test neutrality rather than assume it. The best advice is the one that still makes sense after you have compared the alternatives yourself.