Consulting in the AI Era: Beyond Polished Presentations

Vorontsov came into consulting from custom development, which he still does: clients, he found, need problems solved beyond the code, and they appreciate work done for them before the deal is signed. Fifteen years brought foreign companies — he names Western Digital among them — large Russian holdings from mining to insurance, and foreign subsidiaries of Russian companies: Unitel in Uzbekistan, Skymobile in Kyrgyzstan. He calls himself an adherent of classic business consulting that starts from organisational structure and processes, because a company's options grow out of them. Alexander adds: start from a parochial "let us plug a CRM in here" rather than from the business, and you can execute the project beautifully and still miss the real problem. The guest's recent case is a survey for an import-substitution replacement, where one CRM mixed master data, outlet dashboards, and management reporting produced by rewriting the master base daily.

In his channel's announcement the guest joked that half of all consulting projects are useless because the result is known in advance and merely needs confirming, and the other half because it has to be disputed instead. The value survives in "those very 46%", when the client is ready to change a settled opinion mid-project. In the scenario he walks through on air, replacing an outdated loyalty processing system, the partner the client wanted to lean on could not keep up, and the decision-maker himself proposed other options: custom development, or several small vendors on a single stack. The reverse happens too: a regional company invited him to build a data warehouse, but driving past its billboard he saw a phone number with no call-tracking substitution — the platform was installed, yet no data reached it. A client who runs ahead of the locomotive and has already bought licences he stops gently, borrowing the doctors' practice of not telling a patient how ill he is: work with artefacts and organisational dynamics, pointing out problems already fixed.

The most obvious change is fewer linear and entry-level staff, juniors included, who used to transcribe documents and take notes; juniors have not vanished, but selection is harsher and hires must reach middle-minus in three to six months. A solid task breakdown that once took an architect, an analyst, and a project manager eight hours spread across a week now fits into three hour-long iterations with discussion — and it can be trusted, because it comes from your own thinking while you challenge a half-ready artefact. That toolkit — constraints, risks, assumptions, a calendar plan, effort estimation — his team turned into a product he declined to name, so as not to sound like an advertisement. Projects beyond your own experience have not sped up at all: gamifying a mobile app stitched to a loyalty programme and to assortment reporting down to product roles, where a coffee or a croissant becomes a "cow" or a "star", means joining three layers, and one systems analyst is not enough there.

Alexander describes the same boundary from the other side: as an architect he now does in a day what used to take a day of writing an RFC, a week of explaining it, and a round of approvals — while in an unfamiliar area, such as his home lab for local inference of small models, the bottleneck is not the agent's speed but how fast he can load knowledge into his head, and by evening his cognitive model of the world is overloaded. An expert picks the option like a grandmaster in blitz; a newcomer is left with the old route and a hunt for people carrying the expertise. Vorontsov adds a cognitive trap: people do not distinguish delegating to a person trusted for reputation and results from delegating to an agent they assembled themselves. Alexander answers from the engineering side: hence verification debt — the agent hands you a pull request of fifteen thousand lines, and beyond intent you need acceptance criteria that only a domain expert can write. A closed loop fed with bug fixes holds homeostasis, like an air conditioner or his wife's insulin pump, until a large unpredictable event throws it off the rails.

Fixing someone else's work costs more than building from scratch, and consulting rather benefits from that. A large retailer hired a well-known IT consultancy to design its contact centre's org structure and its move from outsourcing to an internal team, and there the shift schedule was calculated by multiplying call duration by call count — where a competent manager would have used at least an Erlang calculator. Vorontsov did not scold his fellow practitioners; he showed how to plan staff instead: transitions between shifts, and the fifteen-minute reporting without which some conclusions turned out wrong. Alexander recalls learning about Erlang while figuring out a model serving platform. Value keeps moving towards the applied: one client was offered a prototype built straight away on an already validated hypothesis instead of a project worth some sixty million — though that engagement never happened. Consulting, he predicts, will become more applied: delivery turns into part of the consultation itself, as already happened with the Big Four's IT practices, and golden slides in the spirit of a McKinsey deck, Alexander adds, have lost their value.

The sharpest example is a chain of around thirty hotels where some properties are franchises and communication consents are collected by the franchisee: the requirements held neither a consent structure nor consolidation of user records, and the client proposed giving each franchisee its own slice — thirty instances to live with. The consequences of such a decision, the guest says, can be prototyped: three or four hours of having Claude sketch an interface for managing that role model show what administering it would look like and where data would leak — a ready argument against it. In the closing quickfire the guest names the devalued artefact: the list of four or five precise questions that once built trust with a decision-maker. The skill that gained value is the opposite one — writing those questions yourself instead of asking a model. What clients forget to ask about is knowledge transfer: teams are smaller, the expertise sits in three heads, and neither instructions nor a trial operation period carry it across. Hence two pieces of advice: interview the people who will work with their hands rather than the talking head, and agree upfront to take one to three of them onto your payroll.