AI-aided strategy consulting: six legitimate doubts and our answers

What AI-aided strategy consulting is, and what it does better

AI in consulting puts a simple question to clients: does it make the work better, or just faster to produce? Our claim reads: “AI changes not only the pace but consulting itself: facts instead of assumptions, calculated scenarios instead of gut feeling, tools that stay after the project. The judgement calls are made by experienced consultants.” This article tests that claim against six legitimate doubts, each with method, evidence and an honest limit.

Used with discipline, AI makes the work not just faster but better. In the same time, more sources can be reviewed and more options calculated, so you decide on a broader basis. Every number carries its source, every assumption is visible. Instead of a slide deck, you get a calculation model with a cockpit that calculates scenarios at the push of a button and stays in-house after the project. The time saved goes above all to conversations with customers and experts, and to the partners’ judgement.

Doubt 1: Does quality suffer?

The concern about quality is justified, because AI does not improve every task. A field experiment with employees of a management consultancy shows this (Dell’Acqua et al.): on tasks that an AI handles well, participants using AI worked faster and better. On a task where the AI fails, they reached the correct solution less often than without AI. Both occur in the same workflow, often in tasks that look similarly difficult. So quality is built at fixed checkpoints. The honest limit: this takes experienced people’s time. Every AI answer counts as a draft and goes through five checks:

• Confidentiality. Was the content allowed into this tool, and may the result go out in this form?
• Sources. Is every number and every claim backed by the primary source?
• Plausibility. Are the order of magnitude and the logic right, and do independent derivations of key numbers lead to a similar result?
• Status of the hypotheses. Is it clear what is confirmed, what is qualified and what is not supported?
• Accountability. Has a person read, checked and approved the result, and do they put their name to it?

Doubt 2: What happens to our data?

What happens to your data belongs at the start of a project, not at its end. Before the work begins, we agree with you which data may go into which tool. Two gates govern this: “May this go in?” before every input and “May this go out?” before anything is passed on. No AI content leaves the workspace without human review. A traffic light helps with the first gate:

• Green: public. Published figures, studies and websites.
• Amber: properly anonymised. Abstracted content with no names, no real figures and no way of tracing it back to individuals or companies.
• Red: confidential. Real financial and customer data, meeting notes, internal emails: only in a tool your company has approved.
• Black: off limits as a rule. Insider information, HR data and individual personal data, plus anything legally sensitive: never in public applications, and in approved tools only after consultation.

Before the start, also insist on business accounts rather than private logins. Inputs should not be used for training. Clarify where processing takes place and on what contractual basis, and grant only as many access rights as necessary. The German Data Protection Conference also recommends the first two points for AI applications involving personal data (DSK). Your AI policy and your approvals take precedence, and in training exercises only invented or public data is entered into AI tools. The honest limit: some steps therefore run without AI, and that costs speed.

Doubt 3: What if the AI invents numbers or sources?

The risk of invented numbers and sources is real. Language models can produce content that sounds confident but is wrong, up to and including invented reasoning and citations. NIST, a US federal agency, lists such hallucinations as a distinct risk of generative AI and recommends checking sources and citations in AI output (NIST). On top of that comes a tendency to agree: AI assistants often tell users what they want to hear (Sharma et al.). The honest limit: some risk remains, and checking takes time. Five rules help catch such errors:

• AI is not a source. Every number is checked against the primary source and logged in the source register with its citation.
• Keep assumptions separate. Assumptions sit apart from what is evidenced and are labelled as assumptions; the tool, too, is told not to invent numbers.
• Ask for dissent. On instruction, the AI deliberately argues the opposing position; that does not replace a source either.
• Ranges instead of false precision. Where data is thin, a range with its derivation takes the place of a number with decimal places.
• Flag what is open. Provisional findings are labelled provisional, and unchecked ones “to be validated”, until an interview or your own data settles them.

Doubt 4: Can’t anyone do this with ChatGPT and the like?

The honest answer first: a good AI tool can answer many questions on its own, and then you do not need consulting. For a high-stakes strategic question, that is not enough: a chat window answers the question you ask and tends to agree with you. The difference lies in a fixed process in which people check and decide at agreed points. Above all, though, no model delivers these three things:

• The hurdle before the evidence. What an option has to achieve is agreed at the outset, typically in the first week and before any results are in. That way the hurdle cannot be moved afterwards to fit the result.
• Kill criteria. It is fixed in advance which finding ends an option. This is meant to stop mediocre options being carried along to the end.
• Tools that last. A calculation model with a cockpit or a small application needs requirements, tests, documentation and someone to maintain it: more than a chat window delivers.

A RAND interview study of failed AI projects also suggests that the question matters more than the tool. The first of the five main root causes it names is that stakeholders misunderstand or miscommunicate what problem the AI is meant to solve (Ryseff, De Bruhl and Newberry). How a clear question becomes a quantified strategy is the subject of our article on strategy development.

Doubt 5: So what is left for the consultant?

Consultants make the judgement calls and stand behind them. “The AI said so” is no defence: whoever uses a result is accountable for it. At our firm, one more rule applies: two partners, every engagement led personally. The honest limit: what stays with people hardly gets faster with AI, because conversations and consensus take calendar time. That covers five tasks:

• Sharpening the question. People, not models, set the question, the hurdle and the kill criteria.
• Conducting conversations. With customers and experts, on request without revealing your name.
• Weighing contradictions. When pieces of evidence contradict each other, a person decides and explains the choice.
• Writing the synthesis. For each finding: what was heard, which evidence supports it, what follows from it, and whether it is confirmed, qualified or not supported.
• Preparing decisions and handing over. Presenting options to choose from, stating uncertainty openly, and handing over results so that your team can carry on the work.

Doubt 6: Will it be cheaper and faster?

What gets faster is mainly desk research, first drafts and calculating variants. Conversations, checking and decisions hardly do. We do not quote a blanket savings or speed figure for strategy projects. There is no documented before-and-after comparison to support one, and the effects vary widely with the task and with experience. In a field study in customer service, less experienced employees benefited markedly from an AI assistant, while the most experienced and most skilled barely did (Brynjolfsson, Li and Raymond). This does not carry over one-to-one to strategy consulting.

The time saved goes mainly into what makes the difference: checking against primary sources, more conversations with customers and experts, and tools that stay after the project. Because research and drafts take up less effort, projects also become cheaper for our clients; we do not promise a percentage for this. The honest limit: calendar time depends on the state of the data, interviewees’ schedules and decision rounds, not on AI.

Where AI-aided consulting does not help, and when you do not need consulting

Not every question needs consulting, and not every consulting project needs AI. For a simple question with a public answer, a good tool and a quick check of the sources are enough, and we say so. If you want to introduce AI in your own company, our article Introducing AI in your company shows the way, and AI strategy describes our approach. AI-aided consulting achieves little in these situations:

• The decision has already been made. If all you want is confirmation, an AI inclined to agree will give it to you quickly, but without insight.
• There is no data and no access to people. Without your own figures and without conversations, every model remains a collection of assumptions.
• It is mainly about trust, consensus or negotiation. Then conversations, facilitation and time count, not computing power.
• The data is too confidential for any approved tool. Then the work is done without AI, and that is a legitimate decision.