If this is your first time hiring an AI partner, the temptation is to evaluate them the way you evaluate a software vendor: ask about the stack, the certifications, the case studies, the pricing tiers. This is not wrong — but it will not protect you from the most common failure mode, which is hiring someone who is technically competent and operationally hostile to your business.
Here are the six questions that, in our experience, separate partners who deliver from partners who explain.
01. "What did the last engagement look like — week by week?"
A good answer here is unglamorous. Week one was interviews. Week two was a workflow map. Week three was a feasibility note nobody asked for but everyone needed. The bad answer is a sizzle reel.
Ask for a real timeline of a real project. If the partner cannot describe their last engagement in calendar weeks with deliverables attached to each, they do not have a method. They have a brand.
02. "When was the last time you told a client not to build something?"
This question separates consultants who get paid by the hour from partners who care about the outcome. The right answer involves a specific story and an awkward conversation. The wrong answer is "we always find something to build."
Most operator engagements have a build/buy/redesign decision in the first month. A partner who never says "buy" or "redesign first" is selling builds, not solutions.
The most useful sentence we say in a discovery call is often: "this isn't an AI problem."
— a working principle
03. "Who runs the system after you leave?"
This is the question that ends a surprising number of bad engagements before they start. If the answer is "we will, on a retainer," you are renting a system, not buying one. There are good reasons to rent — but you should know that is what you are signing.
The better answer names someone on your team, lays out a handoff plan, and budgets for documentation that survives the partner leaving. The best partners want to be unnecessary by month six.
04. "What does good look like, and how will we measure it?"
If the partner cannot, in fifteen minutes on a discovery call, propose a measurable target — and the harness to measure it — they will not know whether the project worked when it ships.
This is closely related to the eval-first habit we wrote about in the month-three piece. A partner who does not lead with measurement is a partner who will not be able to defend the project when it is questioned later.
05. "Where does our data live during the engagement?"
The answer here should be specific to the level of paranoia your industry requires. For a regulated firm: in your tenant, never copied. For a less regulated one: the partner's environment, with named retention policies. The wrong answer is hand-waving.
This is not just a compliance question. It is a tell about how the partner thinks. Partners who are casual about data are casual about a lot of other things you will only discover later.
06. "Show me a postmortem you wrote for a project that didn't go well."
The answer to this question is the single best signal in the room. A partner who cannot produce one either has never had a project go sideways — unlikely — or does not write postmortems. Either way, they are not a partner who will be honest with you when something goes wrong on your engagement.
The half that isn't about models: questions 01, 02, 03, and 06 do not mention models, frameworks, or benchmarks. They are about how the partner runs themselves. That is what determines whether your project ships.
A short closing
The model is the easy part of any AI engagement. The hard parts are scoping, measurement, ownership, and honesty. Hire for those, and the model usually takes care of itself.
Ako prvi put angažujete AI partnera, iskušenje je da ih ocenjujete kao softverskog dobavljača: pitate za stack, sertifikate, studije slučaja, cenovne nivoe. Nije pogrešno — ali vas to neće zaštititi od najčešćeg načina otkazivanja, a to je angažovanje nekoga ko je tehnički kompetentan i operativno neprijateljski raspoložen prema vašem poslu.
Evo šest pitanja koja, po našem iskustvu, odvajaju partnere koji isporučuju od partnera koji objašnjavaju.
01. „Kako je prošli angažman izgledao — nedelju po nedelju?"
Dobar odgovor ovde nije glamurozan. Prva nedelja su bili intervjui. Druga nedelja je bio mapiranje toka rada. Treća nedelja je bila feasibility beleška koju niko nije tražio, a svima je bila potrebna. Loš odgovor je sizzle rola.
Tražite stvarni vremenski sled stvarnog projekta. Ako partner ne može da opiše prošli angažman u kalendarskim nedeljama sa isporukama vezanim za svaku, nemaju metod. Imaju brend.
02. „Kada ste poslednji put rekli klijentu da ne nešto pravi?"
Ovo pitanje odvaja konsultante koji se plaćaju po satu od partnera kojima je stalo do ishoda. Pravi odgovor uključuje konkretnu priču i neprijatan razgovor. Pogrešan odgovor je „uvek nađemo nešto da napravimo".
Većina operativnih angažmana ima odluku između build/buy/redesign u prvom mesecu. Partner koji nikad ne kaže „kupi" ili „prvo redizajn" prodaje izgradnje, ne rešenja.
Najkorisnija rečenica koju kažemo na razgovoru često je: „ovo nije AI problem."
— radni princip
03. „Ko vodi sistem kad vi odete?"
Ovo pitanje završava iznenađujuće mnogo loših angažmana pre nego što počnu. Ako je odgovor „mi, na retaineru", iznajmljujete sistem, ne kupujete ga. Ima dobrih razloga za iznajmljivanje — ali treba da znate da to potpisujete.
Bolji odgovor imenuje nekoga iz vašeg tima, postavlja plan predaje, i budžetira dokumentaciju koja preživi odlazak partnera. Najbolji partneri žele da budu nepotrebni do šestog meseca.
04. „Kako izgleda dobro, i kako ćemo ga meriti?"
Ako partner ne može, u petnaest minuta na razgovoru, da predloži merljivi cilj — i alat da ga meri — neće znati da li je projekat upalio kada bude isporučen.
Ovo je blisko vezano za naviku „evaluacija prvo" o kojoj smo pisali u tekstu o trećem mesecu. Partner koji ne kreće od merenja je partner koji neće moći da odbrani projekat kada bude doveden u pitanje.
05. „Gde se podaci nalaze tokom angažmana?"
Odgovor ovde treba da bude specifičan za nivo opreznosti koji vaša industrija zahteva. Za regulisanu firmu: u vašem tenant-u, nikad kopirano. Za manje regulisanu: u okruženju partnera, sa imenovanim politikama čuvanja. Pogrešan odgovor je odmahivanje rukama.
Ovo nije samo pitanje usklađenosti. To je signal kako partner razmišlja. Partneri koji su nemarni prema podacima nemarni su i prema mnogim drugim stvarima koje ćete otkriti tek kasnije.
06. „Pokažite mi postmortem projekta koji nije prošao dobro."
Odgovor na ovo pitanje je najbolji pojedinačni signal u prostoriji. Partner koji ne može da ga proizvede ili nikad nije imao projekat koji je krenuo nizbrdo — malo verovatno — ili ne piše postmorteme. U svakom slučaju, nije partner koji će biti iskren prema vama kada nešto pođe naopako u vašem angažmanu.
Polovina koja nije o modelima: pitanja 01, 02, 03 i 06 ne pominju modele, frejmvorke ili benchmarkove. Ona su o tome kako partner vodi sebe. To je ono što određuje da li će vaš projekat biti isporučen.
Kratko zatvaranje
Model je lakši deo svakog AI angažmana. Teški delovi su određivanje opsega, merenje, vlasništvo i iskrenost. Angažujte za to, i model obično vodi sam o sebi računa.
Filed under: STRATEGY · METHOD First published: Mar 19, 2026
