The default UI for an AI feature, in 2026, is a chat box. The default UI for an AI feature in your back office should almost never be a chat box. These two facts have been awkwardly coexisting for two years, and the cost is starting to show up in the projects we audit.
This is a short argument for why chat is the wrong shape for most internal work, and what to use instead.
01. Chat is great for exploration. Bad for repetition.
A chat interface is correctly tuned for the job of figuring out what to ask. Each question is different. The user is exploring. The blank box invites a fresh thought.
Back-office work is the opposite. The user is doing the same thing for the eleventh time today. They know what to ask. They have asked it ten times already. The blank box, in this context, is friction. It is a small but real cognitive tax — a re-articulation of an intent the user has already had.
The teams who hate their AI tools, almost without exception, hate them because they have to retype the same prompt every morning.
The blank chat box is a question the system is asking you. After the third repetition, that's an annoyance, not an interface.
— working note
02. The right interface for repetitive work is a button
Or a form. Or a row in a queue. Or a scheduled job that emits a Slack message. The shape varies, but the pattern is the same: the model's job is to produce a known output for a known shape of input, on a known cadence. The user's job is to handle the exceptions.
This is not a regression. This is the part of the field that the framing of "chatbots" has caused us to forget: that the most valuable AI features in production look almost nothing like the AI features in marketing material. They look like buttons that do useful things, queues that handle drudgery, and notifications that arrive without being asked for.
03. Three back-office patterns that beat chat
In our work, the following patterns reliably outperform a chat interface for internal use:
The "draft and review" queue. The system produces a draft — a quote, a reply, a summary — into the existing tool. The human approves, edits, or rejects. No prompt. No box. Just a list of work that is already partly done.
The "explain this row" overlay. A user opens a record. A side panel summarizes its history, flags anomalies, suggests the next step. The user did not ask. The summary appeared because they navigated to the record.
The scheduled digest. A daily, weekly, or per-event note that arrives in the channel the team already lives in — Slack, email, the existing CRM ticket. No interface to learn. The "interface" is the channel.
None of these are exotic. None of them require a model upgrade. All of them are friction-free in a way that a chat box is not.
04. When chat is right
We are not against chat. We use it daily — for the work it is good at. Chat is right when:
The user does not yet know what to ask.
The work is genuinely one-off.
The model needs clarifying turns to converge on the answer.
Onboarding an analyst into a new dataset, summarizing a long document for a specific purpose, drafting an unusual proposal — these are good chat use cases. The clue is novelty. If the user is doing it for the first time, chat is fine. If they are doing it for the eleventh, it is the wrong shape.
Rule of thumb: if the same prompt is going to be typed by the same person more than three times, it is no longer a prompt. It is a button you have not yet built.
A short closing
Chat is a powerful interface. It is also a lazy default. The teams who get the most value from AI inside their business are the ones who, after the initial chat-everything phase, start asking what is the right interface for this specific job — and then build it.
The blank box is rarely the answer. The button usually is.
Podrazumevani UI za AI funkcionalnost u 2026. godini je čet polje. Podrazumevani UI za AI funkcionalnost u vašem back-officeu skoro nikad ne bi trebalo da bude čet polje. Ove dve činjenice nezgodno koegzistiraju već dve godine, i cena počinje da se vidi u projektima koje pregledamo.
Ovo je kratak argument za to zašto je čet pogrešan oblik za većinu internog rada, i šta koristiti umesto njega.
01. Čet je odličan za istraživanje. Loš za ponavljanje.
Čet interfejs je tačno podešen za posao otkrivanja šta da se pita. Svako pitanje je drugačije. Korisnik istražuje. Prazno polje poziva svežu misao.
Rad u back-officeu je suprotno. Korisnik radi istu stvar po jedanaesti put danas. Zna šta da pita. Pitao je već deset puta. Prazno polje je, u ovom kontekstu, otpor. To je mali ali stvarni kognitivni porez — ponovno artikulisanje namere koju je korisnik već imao.
Timovi koji mrze svoje AI alate, gotovo bez izuzetka, mrze ih jer moraju da pretipkavaju isti prompt svako jutro.
Prazno čet polje je pitanje koje sistem postavlja vama. Posle trećeg ponavljanja, to je nervaža, ne interfejs.
— radna beleška
02. Pravi interfejs za ponavljani rad je dugme
Ili forma. Ili red u redu čekanja. Ili zakazani job koji šalje Slack poruku. Oblik varira, ali obrazac je isti: posao modela je da proizvede poznati izlaz za poznati oblik ulaza, na poznatoj kadenci. Korisnikov posao je da obrađuje izuzetke.
Ovo nije korak unazad. Ovo je deo struke koje smo, uokviravanjem „četbotova", zaboravili: da najvrednije AI funkcionalnosti u produkciji izgledaju gotovo nimalo nalik AI funkcionalnostima u marketinškom materijalu. Izgledaju kao dugmad koja rade korisne stvari, redovi čekanja koji se nose sa sitnim poslovima, i obaveštenja koja stižu bez da se traže.
03. Tri obrasca back-officea koji nadmašuju čet
U našem radu, sledeći obrasci pouzdano nadmašuju čet interfejs za internu upotrebu:
Red „nacrt i pregled". Sistem proizvodi nacrt — ponudu, odgovor, sažetak — u postojećem alatu. Čovek odobrava, uređuje ili odbija. Bez prompta. Bez polja. Samo lista posla koji je delimično već obavljen.
Sloj „objasni ovaj red". Korisnik otvori zapis. Bočni panel sumira njegovu istoriju, označava anomalije, predlaže sledeći korak. Korisnik nije pitao. Sažetak se pojavio jer je stigao do zapisa.
Zakazani rezime. Dnevna, nedeljna ili po događaju beleška koja stiže u kanalu u kome tim već živi — Slack, email, postojeći CRM tiket. Nema interfejsa za učenje. „Interfejs" je kanal.
Nijedan od ovih nije egzotičan. Nijedan ne zahteva nadogradnju modela. Svi su bez otpora na način na koji čet polje nije.
04. Kada jeste čet pravi izbor
Nismo protiv četa. Koristimo ga svakodnevno — za posao za koji je dobar. Čet je pravi izbor kada:
Korisnik još ne zna šta da pita.
Posao je istinski jednokratan.
Model treba pojašnjavajuće runde da konvergira ka odgovoru.
Uvođenje analitičara u novi skup podataka, sumarizacija dugog dokumenta za određenu svrhu, izrada nacrta neuobičajenog predloga — to su dobri slučajevi za čet. Trag je novost. Ako korisnik to radi prvi put, čet je u redu. Ako to radi po jedanaesti put, to je pogrešan oblik.
Pravilo: ako će ista osoba isto kucati isti prompt više od tri puta, to više nije prompt. To je dugme koje još niste izgradili.
Kratko zatvaranje
Čet je moćan interfejs. Ali je i lenj podrazumevani izbor. Timovi koji izvuku najviše vrednosti iz AI-ja u svom poslovanju oni su koji, posle početne čet-sve faze, počnu da pitaju šta je pravi interfejs za ovaj konkretni posao — i onda ga grade.
Prazno polje retko je odgovor. Dugme obično jeste.
Filed under: STRATEGY · UX First published: Feb 08, 2026
