Generative AI consulting: what a consultant does

A generative AI consultant works out where AI genuinely helps your business and then builds it. OpenKit does this in two steps: a diagnosis that maps your goals, workflows, and data before recommending anything, and a build that turns the plan into working systems. The value is not access to a model, which anyone can rent by the month; it is knowing which problem to point it at and being honest when the answer is that you do not need one. This guide covers what that work actually involves, from strategy through the build to proving it paid.

Last updated 10 August 2026.

The term “generative AI” is now so common it risks meaning nothing. It is the headline at every conference and the claimed cure for every business problem, and adoption has climbed fast: McKinsey’s State of AI reporting has tracked organisations moving from cautious pilots to using AI across whole business functions in a short span of years.[1] Behind that rush sits a gap between the hype and a profitable use of it, and the question most leaders are right to ask is how you get from a generic chatbot to something that solves your specific, complicated problem.

The answer is rarely more software. It is knowing which problem is worth solving, and that is the job.

Begin with the problem you are solving

Most AI projects that fail do so before any code is written, because they start with a technology instead of a problem. “We need a chatbot” is not a brief. A generative AI consultant inverts that, beginning with why: what are you actually trying to change, and would AI move it?

Day-to-day operations create a kind of tunnel vision, so the first job is an outside read on the business. That means three things done properly: defining the goal, whether that is cutting cost or shortening a slow process, so the work has something to aim at; mapping the real workflows to find where repetitive, error-prone tasks eat people’s time; and auditing the systems and data, so whatever gets recommended can integrate with what you already run rather than needing a rip-and-replace.

At OpenKit this diagnosis is the opening stage of our AI Audit and Transformation engagement, and it is deliberately builder-neutral: we rank what is worth doing before anyone talks about who builds it. The deliverable is a business document, not a technical one. It sets out the highest-impact opportunities tied to your goals, an honest build-versus-buy call on each, and a sequenced plan with costs, so a leadership team has the clarity to make an investment decision instead of a leap of faith. Our wider AI consulting work sits around that, and the AI consulting guide is a useful companion if you are weighing whether to bring a partner in at all.

What a consultant actually builds

Once the strategy is clear, the work shifts to building, and it tends to fall into three kinds of value. None of them is exotic; the skill is in choosing the right one for the problem and making it hold up in production.

Automating the work nobody should be doing by hand

The most immediate use of generative AI is intelligent automation: replacing manual, repetitive tasks with AI agents so people can spend their time on work that needs judgement. This matters most because it breaks a relationship every growing business hits, where more customers mean more tickets mean more hires. Automation lets you handle a lot more volume without the headcount rising in lockstep, so instead of thirty support agents you might run a small team backed by a capable system.

In practice this shows up across departments. Customer-service assistants handle common questions around the clock and escalate the hard ones to a person, which cuts waiting times without cutting quality. Marketing and sales teams use AI to draft first versions of posts, emails, and outreach, then edit rather than start from a blank page. And core operations get lighter when a system reads invoices and contracts and pulls out the key fields, turning hours of data entry into quick validation. The pattern underneath all of it is the same: the machine takes the repeatable volume and a person owns anything that carries risk.

Turning scattered documents into answers

Every organisation sits on a large, mostly untapped asset: the knowledge locked in PDFs, emails, chat logs, and transcripts that pile up daily. It is fragmented, siloed, and effectively unsearchable, so the same questions get asked and answered again and again, and when an experienced person leaves, their knowledge tends to leave with them.

A consultant’s job here is to turn that mess into a secure, searchable knowledge base that your team can actually query. This is core work for us, and OpenKit’s BAiSICS platform is a working example: it applies AI to legal and commercial document review, taking a review that once ran to around two hours down to roughly ten minutes, at 92% faster processing and 96% extraction accuracy, and saving the firm over £200,000 a year. The reason to build your own rather than lean on a public tool is that a public model knows the open internet and nothing about your confidential work; a private knowledge engine becomes an advantage a competitor cannot copy.

The mechanics, in plain terms, are straightforward. The system connects securely to where your documents live and reads them, converting their meaning into a form it can search. That search works on meaning rather than exact keywords, so a question phrased in normal language finds the right passage. And a simple interface lets someone ask in plain English and get a concise, cited answer drawn from your own documents. That last part, grounding answers in your trusted data, is retrieval-augmented generation. It cuts down the inventing that a public model does when it is asked about material it has never seen, and because every answer carries its source, the reader can check the ones that matter rather than trusting the tone.

Co-pilots that make experts faster

The loud version of the AI story is about replacing people. The useful version is about augmenting them. The strongest tools act as co-pilots that take a specific, high-friction task off an expert’s plate so they can do more of the work only they can do, and by removing the drudgery they tend to make the job better rather than more precarious.

This lands hardest in document-heavy fields. In legal and property work, review and due diligence are slow and mentally taxing, which is exactly where a tool like BAiSICS earns its place, and it is used by chartered surveyors and legal teams to accelerate commercial-lease review. In insurance, a co-pilot can read policy documents, historical claims, and risk reports to hand an underwriter a summary in seconds, and can triage routine claims while flagging the odd ones for a human. In private equity, an analyst can point a system at a data room of thousands of documents and ask, in plain English, for every change-of-control clause across the target’s contracts, and get a synthesised answer in minutes rather than reading for a fortnight. In each case the expert stays in charge; the tool just clears the path.

Proving it paid

A credible consultant ties every recommendation to measurable value, because “we did some AI” is not a result. Measuring return is not complicated in principle. You count the full investment, the build, the infrastructure, the training, and the ongoing running, and you weigh it against the value created, which is a mix of hard savings and slower-burning gains like better decisions, higher retention, and lower churn.

The honest part is admitting which of those you can measure cleanly and which you are estimating. Time saved on a high-volume task multiplied by a loaded cost is concrete. A lift in customer satisfaction is real but harder to bank. A good partner is straight about the difference and does not dress an estimate up as a fact. Because the return question deserves its own treatment, we cover the calculation in detail in the business case for AI guide.

The way to move from estimate to evidence is to prove the idea on one workflow before committing to a large build. That is what the AI Audit and Transformation engagement is for: it is fixed scope and fixed fee, it targets your most valuable problem first, and it produces a real-world result you can build the full case on, rather than a projection. Each step prices the next, so you are never buying further than you can see.

Choosing the right partner

The success of an AI initiative rests heavily on who you pick, and the useful distinction is between a vendor selling a product and a partner solving your problem. A short checklist separates them.

Ask to see real work, ideally in your field. A slick deck proves nothing; a portfolio of shipped projects proves a lot. Ours runs from the Department for Education marking platform through to legal document systems, which is the range you want to see.

Insist on direct access to the people building your solution, not a wall of account managers between you and the engineers. Be wary of anyone pushing a specific product before they have understood your problem, because a real partner starts with a diagnosis, not a demo. Check that support does not end at launch, since AI systems drift and need monitoring and updates. And listen to whether they speak your language: a good consultant is as comfortable talking about cost, risk, and outcomes as about models, and frames the work around your business rather than their technology.

A generative AI consultant, done well, is equal parts strategist and engineer: someone who aligns the work with what matters and builds something that holds up, without losing sight of what each pound is actually buying. The goal was never to “do AI”. It is to use it where it genuinely helps and to leave it alone where it does not.

References

1. McKinsey & Company. (2025). The State of AI: how organizations are rewiring to capture value, accessed on 8 July 2026, https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai