Wɔ saa nsɛm a ɛtoatoa so yi fã a edi kan no mu no, yɛde nsakrae titiriw a efi awo mu kɔ awofo nyansa a wɔde ayɛ nneɛma so sii hɔ. Yɛhwehwɛɛ nea enti a saa ahurututu yi a ɛfiri nyansahyɛ mu kɔ agodie mu yi hwehwɛ sɛ wonya adwene ne akwan ho adwinnadeɛ foforɔ ma UX nhwehwɛmufoɔ, nneɛma so ahwɛfoɔ, ne akannifoɔ. Yɛkyerɛkyerɛɛ agentic suban ahorow a wɔahyehyɛ mu, efi nyansahyɛ so kosi sɛ wɔbɛyɛ ade wɔ ahofadi mu, yɛkyerɛkyerɛɛ nhwehwɛmu akwan a ɛho hia no mu, kyerɛkyerɛɛ asiane ahorow a ɛwɔ agent sludge mu, na yɛde akontaabu metrics a ɛho hia na ama wɔatumi akɔ asasesin foforo yi mu no sii hɔ. Yɛkaa nea enti a wɔyɛɛ saa no ho asɛm. Afei, yɛfiri fapem so kɔ dwumadie so. Saa asɛm yi de ɔkwan a wɔfa so yɛ no ma: nhyehyeɛ a ɛyɛ nokware, dwumadie nhyehyɛeɛ, ne ahyehyɛdeɛ mu nneyɛeɛ a ɛho hia ma agyapadeɛ nhyehyɛeɛ a ɛnyɛ sɛ ɛwɔ tumi nko na mmom ɛyɛ nea ɛda adi, ɛtumi di so, na ɛfata sɛ wɔde di dwuma no mu ahotosoɔ nso. Sɛ yɛn nhwehwɛmu no ne adwinnade a wɔde hu yare a, saa nhwɛso ahorow yi ne ayaresa nhyehyɛe. Wɔyɛ akwan a mfaso wɔ so a yebetumi afa so ama wɔn a wɔde di dwuma no anya tumi a wotumi hu, bere mpo a yɛma AI ahofadi a ebi mmae da no. Botae no ne sɛ wɔbɛbɔ osuahu bi a ahofadi te nka sɛ hokwan a nea ɔde di dwuma no de ama, na ɛnyɛ hokwan a nhyehyɛe no agye. Core UX Nhwɛsode Ma Agent Systems Designing for agentic AI yɛ nhyehyɛe ma abusuabɔ. Ɛsɛ sɛ wɔde saa abusuabɔ yi, te sɛ fekubɔ biara a edi mu no, si nkitahodi a emu da hɔ, wɔn ho wɔn ho ntease, ne ahye a wɔde asi hɔ so. Sɛ yɛbɛdi nsakraeɛ a ɛfiri nyansahyɛ mu kɔ adeyɛ mu no ho dwuma a, yɛde nhwɛsoɔ nsia a ɛdi dwumayɛ asetena a ɛwɔ ananmusifoɔ nkitahodiɛ mu akyi di dwuma:
Pre-Action (Establishing Intent)Intent Preview ne Autonomy Dial no hwɛ hu sɛ nea ɔde di dwuma no kyerɛkyerɛ nhyehyɛe no ne agent no hye mu ansa na biribiara asi. In-Action (Providing Context)The Explainable Rationale and Confidence Signal no kɔ so yɛ nea ɛda adi pefee bere a ɔnanmusifo no reyɛ adwuma, na ɛkyerɛ “nea enti a” ne “sɛnea ɛyɛ nokware.” Post-Action (Safety and Recovery)Action Audit & Undo ne Escalation Pathway no ma ahobanbɔ afiri ma mfomsoɔ anaa mmerɛ a emu nna hɔ kɛseɛ.
Wɔ aseɛ ha yi, yɛbɛka nhwɛsoɔ biara ho asɛm kɔ akyiri, a nyansahyɛ ahodoɔ a ɛfa metrics a ɛbɛma asi yie ho ka ho. Saa botaeɛ yi yɛ ananmusifoɔ nsusuiɛ a egyina nnwuma gyinapɛn so; siesie wɔn gyina wo domain asiane pɔtee no so. 1. Adwene no Nhwɛso: Dɛn ne Ɔkwan a Wɔfa so Yɛ no a Wɔbɛma Ada hɔ Saa nhwɛso yi ne nkɔmmɔbɔ mu ade a ɛne sɛ yɛbɛka sɛ, “Nea merebɛyɛ ni. So wo ho ye wɔ saa asɛm no ho?” Ɛyɛ fapem bere a wɔde hwehwɛ pene wɔ ɔdefo ne ɔnanmusifo abusuabɔ mu. Ansa na ɔnanmusifo bi bɛyɛ ade titiriw biara no, ɛsɛ sɛ nea ɔde di dwuma no nya nea ɛrebɛba no ho ntease a emu da hɔ a emu nna hɔ. Intent Preview, anaa Plan Summary no de pene a wɔde nimdeɛ agye atom si hɔ. Ɛyɛ nkɔmmɔbɔ mu home a wɔde gyina hɔ ansa na wɔayɛ ade, a ɛdannan adaka tuntum bi a ɛwɔ akwan horow a ɛyɛ ne ho mu ma ɛbɛyɛ nhyehyɛe a ɛda adi pefee, a wotumi hwɛ mu. Adwene mu NhyɛaseSɛ wɔde nhyehyɛe bi ma ansa na wɔayɛ ade a, ɛtew adwene mu adesoa so na eyi ahodwiriw fi hɔ, na ɛma wɔn a wɔde di dwuma no nya bere tiaa bi de hwɛ sɛ ɔnanmusifo no te wɔn adwene ase ampa. Anatomy of an Effective Intent Nhwɛso a Wɔde Di Kan:
Nea emu da hɔ na ɛyɛ tiawaƐsɛ sɛ nea wɔadi kan ahwɛ no yɛ nea wotumi di ntɛm ara. Ɛsɛ sɛ ɛbɔ nneyɛe atitiriw ne nea efi mu ba no mua wɔ kasa a emu da hɔ mu, na ɛkwati mfiridwuma mu kasafĩ. Sɛ nhwɛso no, sɛ́ anka ɛbɛyɛ “Executing API call to cancel_booking(id: 4A7B),” ɛsɛ sɛ ɛka sɛ, “Cancel flight AA123 to San Francisco.” Anamɔn a Ɛtoatoa soWɔ anammɔn pii dwumadie mu no, ɛsɛ sɛ nhwɛsoɔ no kyerɛ akwan titire no. Eyi da agent’s logic adi na ɛma wɔn a wɔde di dwuma no hu nsɛm a ebetumi aba wɔ nnidiso nnidiso a wɔahyɛ ho nyansa no mu. Popa Ɔdefoɔ NneyɛeɛNhwɛsoɔ no yɛ gyinaesie, ɛnyɛ amanneɛbɔ kɛkɛ. Ɛsɛ sɛ wɔde nneɛma a wobetumi apaw a emu da hɔ ka ho. Ɛyɛ bere a wɔahyɛ da ayɛ akasakasa, ‘ahoɔhare a ɛkɔ soro’ wɔ adeyɛ no mu a wɔayɛ sɛ wɔde bɛhwɛ sɛ nea ɔde di dwuma no repaw nea ɔde nhumu, titiriw ma nneyɛe a wontumi nsakra anaasɛ ɛyɛ nea ɛho hia kɛse.
Momma yɛnsan nhwɛ yɛn akwantu boafo tebea a efi nsɛm a ɛtoatoa so yi fã a edi kan no mu. Yɛde saa ɔboafo a ɔyɛ nnam yi di dwuma de kyerɛ sɛnea ɔnanmusifo bi di wimhyɛn a wɔtwa mu ho dwuma. Ɔnanmusifo no ahu sɛ wɔatwa wimhyɛn mu na wayɛ nhyehyɛe bi a ɛbɛma wɔasan anya ahoɔden. Intent Preview no bɛyɛ sɛ nea ɛte sɛɛ: Proposed Plan for Your Trip DisruptionM’ahu sɛ wɔatwa wo 10:05 AM wimhyɛn no mu. Here’s what I plan to do:Cancel Flight UA456Process refund and confirm cancellation details.Rebook on Flight DL789Book a confirmed seat on a 2:30 PM wimhyɛn a enni gyinabea, efisɛ eyi ne wimhyɛn a edi hɔ a ɛwɔ hɔ a enni gyinabea a ɛwɔ aconfirmed seat.Update Hotel ReservationMa Marriott amanneɛ sɛ wobɛba akyiri.Email Updated ItineraryFa wimhyɛn foforo ne ahɔhodan ho nsɛm kɔma wo ne wo boafo, Jane Doe.[ Kɔ so yɛ Nhyehyɛe yi ] [ Sesa Nhyehyɛe ] [ M’ankasa Di ho dwuma ]
Saa nhwɛsoɔ yi yɛ adwuma yie ɛfiri sɛ ɛma mfonini a ɛdi mũ, ɛfiri sɛ wɔtwa mu kɔsi nkitahodie so, na ɛma akwan soronko mmiɛnsa a wɔfa so kɔ w’anim: pene a wɔpene so koraa (Toa so), ɔpɛ a wɔwɔ sɛ wɔbɛsesa (Sesa Nhyehyɛeɛ), anaasɛ wɔbɛpopa koraa (M’ankasa di ho dwuma). Saa tumidi a ɛwɔ afã horow pii yi ne ahotoso nnyinaso.
Bere a Ɛsɛ sɛ Wɔde Saa Nhwɛso Yi Di Kan Saa nhyehyɛe yi nyɛ nea wontumi nsusuw ho mma adeyɛ biara a wontumi nsakra (e.g., nea ɔde di dwuma no data a wobɛpopa), ɛfa sikasɛm mu nkitahodi a ɛyɛ sika biara ho, ɛne nnipa anaa nhyehyɛe afoforo kyɛ nsɛm, anaasɛ ɛyɛ nsakrae kɛse a ɛnyɛ den sɛ ɔdefo bi ntumi nsan nhyɛ mu. Asiane a Ɛwɔ OmissionSɛ eyi nni hɔ a, wɔn a wɔde di dwuma no te nka sɛ agent no nneyɛe abɔ wɔn ho ban na wɔbɛma feature no ayɛ adwuma na ama wɔasan adi so. Metrics a Ɛma Odi Yiye:
Gye a Wɔgye tom NsusuiɛNhyehyɛeɛ a Wɔgye tom a Wɔansiesie / Wɔada Nhyehyɛeɛ a Wɔagye Nyinaa adi. Botae a wɔde asi wɔn ani so > 85%. Override FrequencyTotal Di ho dwuma M’ankasa Klik / Total Nhyehyɛe a Wɔada no adi. Rate > 10% kanyan model nhwehwɛmu. Kae PɛpɛɛpɛƆha biara mu nkyekyɛmu a wɔn a wɔde wɔn ho hyɛɛ sɔhwɛ no mu a wobetumi akyerɛw nhyehyɛe no anammɔn yiye wɔ sikani 10 akyi wɔ sikɔne 10 akyi a wɔde nhwɛso no asie.
Eyi a Wɔde Di Dwuma Wɔ High-Stakes Domains mu Bere a akwantu nhyehyɛe yɛ mfitiasede a wotumi ka ho asɛm no, saa nhyehyɛe yi bɛyɛ nea ɛho nhia wɔ mmeae a ɛyɛ den, a asiane kɛse wom a mfomso bi de nea ɛboro ɔhaw ba ankorankoro bi a ɔretu kwan so. Yɛn mu pii yɛ adwuma wɔ mmeae a gyinaesi a ɛnteɛ betumi ama nhyehyɛe bi agyae, de ɔyarefo bi ahobammɔ ato asiane mu, anaasɛ asiane afoforo pii a mfiridwuma a wontumi mfa ho nto so de bɛba no mu. Susuw DevOps Release Agent a wɔde ahyɛ ne nsa sɛ ɔnhwɛ cloud infrastructure so ho. Wɔ saa tebea yi mu no, Intent Preview no yɛ adwuma sɛ ahobammɔ akwanside a ɛko tia bere a wɔde gyae adwuma wɔ akwanhyia mu.
Wɔ saa nkitahodi yi mu no, nsɛmfua pɔtee (Drain Traffic, Rollback) no si generalities ananmu, na nneyɛe no yɛ abien na ɛwɔ nkɛntɛnso. Ɔdefoɔ no ma kwan ma wɔyɛ nsakraeɛ kɛseɛ wɔ adwumayɛ mu a egyina agent’s logic so, sene sɛ ɔbɛpene nyansahyɛ bi so. 2. The Autonomy Dial: Calibrating Trust ne Nkɔso Tumi Abusuabɔ biara a ɛfata wɔ anohyeto ahorow. Autonomy Dial no yɛ sɛdeɛ ɔdefoɔ no de si hɔ ne wɔn agent, kyerɛkyerɛ deɛ wɔn ho tɔ wɔn sɛ agent no ankasa bedi ho dwuma. Ahotoso nyɛ nsakrae a ɛba abien; ɛyɛ spectrum a ɛyɛ spektrum. Ebia obi a ɔde di dwuma benya ɔnanmusifo bi mu ahotoso sɛ obedi nnwuma a ɛho nhia pii ho dwuma wɔ ne ho nanso ahwehwɛ sɛ wosi so dua koraa wɔ gyinaesi ahorow a ɛho nhia pii ho. Autonomy Dial, a ɛyɛ tumi krataa a ɛkɔ so nkakrankakra no ma wɔn a wɔde di dwuma no tumi de wɔn a wɔpɛ sɛ wɔyɛ agent ahofadi no si hɔ, na ɛma wɔyɛ nnipa a wɔde wɔn ho hyɛ abusuabɔ no mu denneennen. Psychological UnderpinningSɛ wɔma wɔn a wɔde di dwuma no kwan ma wɔyɛ agent no ahofadi no ho nhyehyɛe a, ɛma wonya beae a wobetumi adi so, na ɛma wɔde nhyehyɛe no nneyɛe ne wɔn ankasa asiane a wotumi gyina ano no hyia. DwumadieEyi betumi adi dwuma sɛ nhyehyeɛ a ɛnyɛ den, ɛda adi pefee wɔ aplikeshɔn no mu, sɛdeɛ ɛbɛyɛ yie no, wɔ adwuma biara mu. Sɛ yɛde taxonomy a efi yɛn asɛm a edi kan no mu di dwuma a, nhyehyɛe ahorow no betumi ayɛ:
Hwɛ & Fa nyansahyɛ Mepɛ sɛ wɔbɔ me amanneɛ wɔ hokwan anaa nsɛm bi ho, nanso ɔnanmusifo no renhyɛ nhyehyɛe bi ho nyansa da. Plan & ProposeAgent no betumi ayɛ nhyehyɛe, nanso ɛsɛ sɛ mesan hwɛ emu biara mu ansa na wɔayɛ ade biara. Fa Confirmation yɛ adeMa nnwuma a wonim no yiye no, agent no betumi asiesie nneyɛe, na mɛma kɔ/no-go confirmation a etwa to. Yɛ Adwuma wɔ Ne Ho Wɔ nnwuma a wɔadi kan apene so (e.g., akasakasa a wɔbɔ wɔ sika a ɛba fam sen $50) ho no, ɔnanmusifo no betumi ayɛ ade wɔ ne ho na wabɔ me amanneɛ wɔ nokwasɛm no akyi.
Sɛ nhwɛso no, email boafo betumi anya autonomy dial a ɛyɛ soronko a wɔde yɛ nhyiam ho nhyehyɛe sen sɛ wɔde email bɛmena wɔ nea ɔde di dwuma no ananmu. Saa granularity yi yɛ ade titiriw, efisɛ ɛda nokwasɛm a ɛyɛ nuanced a ɛwɔ obi a ɔde di dwuma no mu ahotoso adi. Bere a Ɛsɛ sɛ Wɔde Saa Nhwɛso Yi Di Kan Fa eyi di kan wɔ nhyehyɛe ahorow a nnwuma gu ahorow kɛse wɔ asiane ne ankorankoro pɛ mu (e.g., sikasɛm nhyehyɛe nnwinnade, nkitahodi nhyehyɛe ahorow). Ɛho hia ma onboarding, ɛma wɔn a wɔde di dwuma no tumi fi ase de ahofadi a ɛba fam na ɛma ɛkɔ soro bere a wɔn ahotoso kɔ soro no. Asiane a Ɛwɔ Omission muSɛ eyi nni hɔ a, wɔn a wɔde di dwuma a wohu huammɔdi biako no begyae agent no koraa sen sɛ wɔbɛsan afrɛ ne tumi krataa no kɛkɛ. Metrics a Ɛma Odi Yiye:
Trust DensityƆha biara mu nkyekyɛmu a wɔde di dwuma wɔ nhyehyɛe biara mu (e.g., 20% Fa nyansahyɛ, 50% Si so dua, 30% Auto). Setting ChurnNsiesiei Nsakraeɛ dodoɔ / Nnipa a wɔyɛ adwuma nyinaa bosome biara. High churn kyerɛ ahotosonneɛma a ɛsakrasakra.
3. Ntease a Yebetumi Akyerɛkyerɛ Mu: Dɛn Nti na Wobebua? Bere a ɔhokafo pa ayɛ ade bi akyi no, ɔkyerɛkyerɛ wɔn nsusuwii mu. Saa nhwɛsoɔ yi ne nkitahodi a ɛda adi pefee a ɛdi adeyɛ bi akyi, a ɛbua Dɛn ntia? ansa na wɔabisa mpo. “Meyɛɛ saa efisɛ woaka akyerɛ me bere bi a atwam sɛ wopɛ X.” Sɛ ɔnanmusifo bi yɛ ade, titiriw wɔ ne ho a, asɛmmisa a ɛba nea ɔde di dwuma no adwene mu ntɛm ara no taa yɛ sɛ, Dɛn nti na ɛyɛɛ saa? Explainable Rationale nhyehyɛe no de nnam bua asɛmmisa yi, na ɛde ntease tiawa ma wɔ agent’s gyinaesi ahorow ho. Eyi nyɛ mfiridwuma mu log fael. Wɔ m’asɛm a edi kan wɔ saa nsɛm a ɛtoatoa so yi mu no, yɛkaa nhyehyɛe a edi kan a yɛbɛkyerɛ ase akɔ kasa a nea ɔde di dwuma no anim de asiw nnaadaa ano ho asɛm. Saa nhwɛso yi ne saa nnyinasosɛm no a wɔde di dwuma wɔ ɔkwan a mfaso wɔ so so. Ɛdan raw logic no ma ɛbɛyɛ nkyerɛkyerɛmu a onipa betumi akenkan a egyina nea ɔde di dwuma no ankasa apɛde a waka ne nsɛm a wadi kan de ahyɛ mu so. Adwene mu NhyɛaseSɛ wotumi kyerɛkyerɛ agent bi nneyɛe mu a, ɛte nka sɛ ntease wom sen sɛ ɛbɛyɛ nea ɛba kwa, na ɛboa nea ɔde di dwuma no ma ɔyɛ adwene mu nhwɛso a edi mu a ɛkyerɛ sɛnea agent no susuw. Ntease ahorow a Etu mpɔn:
Grounded in PrecedentNkyerɛkyerɛmu a eye sen biara no san kɔ mmara, nea wɔpɛ, anaa adeyɛ bi a wɔadi kan ayɛ so. Simple ne DirectKwati tebea mu ntease a ɛyɛ den. Fa “Esiane sɛ wokaa X nti, meyɛɛ Y” nhyehyɛe a ɛnyɛ den di dwuma.
Sɛ yɛsan kɔ akwantuo nhwɛsoɔ no so a, sɛ wɔsan kyerɛw wimhyɛn no wɔ ne ho akyi a, ebia nea ɔde di dwuma no behu eyi wɔ wɔn amanneɛbɔ feed no mu: I’ve rebooked your canceled flight.New Flight: Delta 789, departing at 2:30 PM.Nea enti a meyɛɛ saa adeyɛ yi:Wimhyɛn adwumakuo no twaa wo mfitiaseɛ wimhyɛn no mu.Woadi kan apene autonomous rebooking for same-day, non-stop flights.[ View New Itinerary ] [ Undo this Action ]
Ntease a ɛwɔ mu no mu da hɔ, wotumi bɔ ho ban, na ɛhyɛ adwene a ɛne sɛ ɔnanmusifo no reyɛ adwuma wɔ ahye a nea ɔde di dwuma no de sii hɔ no mu den. Bere a Ɛsɛ sɛ Wɔde Saa Nhwɛso Yi Di kan Fa no di kan ma adeyɛ biara a ɛyɛ ne ho a ntease no nna adi ntɛm ara fi nsɛm a ɛfa ho no mu, titiriw ma nneyɛe a ɛkɔ so wɔ akyi anaasɛ abɔnten so adeyɛ bi na ɛkanyan no (te sɛ wimhyɛn a wɔtwa mu nhwɛso no). Asiane a Ɛwɔ sɛ Wɔayi Afiri MuSɛ eyi nni hɔ a, wɔn a wɔde di dwuma no kyerɛ nneyɛe a ɛfata a wɔde wɔn ho hyɛ mu ase sɛ nneyɛe a wɔanhyɛ da anaasɛ ‘mmoe,’ na ɛmma wontumi nhyehyɛ adwene mu nhwɛso a ɛteɛ. Metrics a Ɛma Odi Yiye:
Dɛn ntia? Ticket VolumeMmoa tekiti dodow a wɔakyerɛw so sɛ “Agent Behavior — Unclear” wɔ nnipa 1,000 a wɔyɛ nnam biara mu. Rationale ValidationƆha biara mu nkyekyɛmu a wɔde di dwuma a wɔde nkyerɛkyerɛmu no toto ‘Ɛboa’ wɔ nkitahodi akyi microsurveys mu.
4. Ahotoso Nsɛnkyerɛnne no Saa nhwɛso yi fa agent no a ɔnim ne ho wɔ abusuabɔ no mu ho. Ɛdenam n’ankasa ahotoso a ɛka ho no so no, ɛboa nea ɔde di dwuma no ma osi bere a ɛsɛ sɛ ɔde ne ho to n’atemmu so ne bere a ɔde nhwehwɛmu pii bedi dwuma ho gyinae. Sɛnea ɛbɛyɛ na wɔaboa wɔn a wɔde di dwuma no ma wɔasusuw wɔn ankasa ahotoso ho no, ɛsɛ sɛ ɔnanmusifo no da n’ankasa ahotoso adi wɔ ne nhyehyɛe ne ne nneyɛe mu. Eyi ma agent no mu tebea yɛ nea wotumi kenkan yiye na ɛboa nea ɔde di dwuma no ma osi bere a ɛsɛ sɛ ɔhwehwɛ gyinaesi bi mu yiye ho gyinae. Adwene mu NhyɛsoSurfacing uncertainty boa ma wosiw automation bias ano, ɛhyɛ wɔn a wɔde di dwuma no nkuran sɛ wɔnhwehwɛ nhyehyɛe ahorow a ahotoso kakraa bi na ɛwɔ mu sen sɛ wobegye atom anifurae so. Sɛnea wɔde di dwuma:
Ahotosoɔ NkonimdieƆha mu nkyekyɛmu a ɛnyɛ den (e.g., Ahotosoɔ: 95%) betumi ayɛ nsɛnkyerɛnneɛ a ɛyɛ ntɛm, a wɔtumi scan. Scope DeclarationAsɛm a emu da hɔ a ɛfa agent’s area of expertise (e.g., Scope: Travel bookings nkutoo) boa ma wɔhwɛ nea ɔde di dwuma no akwanhwɛ so na esiw wɔn kwan sɛ wɔbɛka akyerɛ agent no sɛ ɔnyɛ nnwuma a wɔannwene no amma. Nsɛnkyerɛnne a Wɔde Aniwa HuAhyɛnsode a ɛyɛ ahabammono betumi akyerɛ ahotoso kɛse, bere a asɛmmisa agyiraehyɛde kɔkɔɔ betumi akyerɛ sɛ wontumi nsi pi, na akanyan nea ɔde di dwuma no ma wahwɛ mu yiye.
Bere a ɛsɛ sɛ wode saa nhyehyɛe yi di kan fa bere a agent no adwumayɛ betumi ayɛ soronko kɛse a egyina input data no su anaa adwuma no mu nsɛm a emu nna hɔ so. Ɛsom bo titiriw wɔ animdefo nhyehyɛe ahorow mu (e.g., aduruyɛ mmoa, mmara aboafo) a ɛsɛ sɛ onipa hwehwɛ AI’s output no mu yiye. Asiane a Ɛwɔ Sɛ Wɔayi Afi MuSɛ eyi nni hɔ a, wɔn a wɔde di dwuma no bɛhwe ase wɔ automation bias mu, anifurae so agye adwenem naayɛ a wonni ahotoso kɛse atom, anaasɛ wɔde ahoyeraw bɛhwɛ adwuma a ahotoso kɛse wom mu mprenu. Metrics a Ɛma Odi Yiye:
Calibration ScorePearson abusuabɔ a ɛda Model Confidence Score ne User Acceptance Rate ntam. Botae a wɔde asie > 0.8. Scrutiny DeltaNsonsonoe a ɛda bere a wɔde hwɛ nhyehyɛe a ahotoso a ɛba fam ne nhyehyɛe a ahotoso kɛse wom ntam. Wɔhwɛ kwan sɛ ɛbɛyɛ papa (e.g., +12 seconds).
5. Adeyɛ no Audit & Undo: Ahobammɔ a Etwa To Ahotoso hwehwɛ sɛ wuhu sɛ wubetumi anya ahoɔden afi mfomso bi mu. Na Undo nodwumadi ne abusuabɔ ahobammɔ afiri a etwa to, a ɛma nea ɔde di dwuma no awerɛhyem sɛ sɛ ɔnanmusifo no nte ase yiye mpo a, nea efi mu ba no nyɛ ɔsɛe. Adwinnade biako pɛ a tumi wom sen biara a wɔfa so ma wɔn a wɔde di dwuma no nya ahotoso ne tumi a ɛyɛ mmerɛw sɛ wɔbɛdan agent bi adeyɛ. Action Audit log a ɛkɔ so tra hɔ, a ɛnyɛ den sɛ wobɛkenkan, a Undo button a ɛda adi pefee ma adeyɛ biara a ebetumi aba no ne ahobammɔ afiri a etwa to. Ɛtew asiane a wosusuw sɛ ɛwɔ ahofadi a wɔbɛma mu no so kɛse. Adwene mu NhyɛsoSɛ́ wunim sɛ ɛnyɛ den sɛ wobetumi asiesie mfomso bi no de adwene mu ahobammɔ ba, na ɛhyɛ wɔn a wɔde di dwuma no nkuran sɛ wɔmfa nnwuma bɛhyɛ wɔn nsa a wonsuro sɛ nea ebefi mu aba a wontumi nsakra. Nneyɛe a Ɛyɛ Paara a Wɔyɛ: Design:
Bere nhyehyɛe HwɛMmere nhyehyɛe kyerɛwtohɔ a ɛfa nneyɛe a agent-initiated nyinaa ho ne ɔkwan a ɛyɛ mmerɛw sen biara. Clear Status IndicatorsKye sɛ adeyɛ bi dii nkonim, ɛrekɔ so, anaasɛ wɔasan ayɛ. Bere a Wɔayɛ no UndosWɔ nneyɛe a ɛbɛyɛ nea wontumi nsakra wɔ bere pɔtee bi akyi (e.g., booking a wontumi nsan mfa mma) no, ɛsɛ sɛ UI no di nkitaho pefee wɔ saa bere mfɛnsere yi mu (e.g., Undo a ɛwɔ hɔ simma 15). Saa pefeeyɛ yi a ɛfa nhyehyɛe no anohyeto ahorow ho no ho hia te sɛ undo tumi no ankasa. Sɛ wodi nokware wɔ bere a adeyɛ bi bɛyɛ nea ɛtra hɔ daa ho a, ɛma wonya ahotoso.
Bere a Ɛsɛ sɛ Wɔde Saa Nhwɛso Yi Di KanEyi yɛ fapem nhyehyɛe a ɛkame ayɛ sɛ ɛsɛ sɛ wɔde di dwuma wɔ agentic nhyehyɛe nyinaa mu. Ɛnyɛ nea wontumi nsusuw ho koraa bere a wɔde nneɛma a ɛma obi tumi yɛ adwuma reba anaasɛ bere a mfomso bi (sikasɛm, asetra, anaa data ho ka) ho ka yɛ kɛse no. Asiane a Ɛwɔ sɛ Wɔayi Afiri MuSɛ eyi nni hɔ a, mfomso biako sɛe ahotoso koraa, efisɛ wɔn a wɔde di dwuma no hu sɛ wonni ahobammɔ afiri biara. Metrics a Ɛma Odi Yiye:
Reversion RateUndone Nneyɛe / Nneyɛe a Wɔayɛ nyinaa. Sɛ Reversion Rate > 5% ma adwuma pɔtee bi a, gyae automation ma saa adwuma no. Safety Net ConversionƆha biara mu nkyekyɛmu a wɔde di dwuma a wɔyɛ upgrade kɔ Act Autonomously wɔ nnafua 7 akyi a wɔde Undo adi dwuma yie.
6. Escalation Pathway: Sɛnea Wodi Nneɛma a Wontumi nsi pi ho dwuma wɔ ɔkwan a ɛyɛ fɛ so Ɔhokafo a onim nyansa nim bere a ɛsɛ sɛ ɔsrɛ mmoa mmom sen sɛ obesusuw ho. Saa nhyehyɛe yi ma ɔnanmusifo no tumi di nsɛm a emu nna hɔ ho dwuma fɛfɛɛfɛ denam kɔ soro a ɔbɛkɔ so akɔ nea ɔde di dwuma no nkyɛn, na ɔda ahobrɛase a ɛma ahotoso ba, sen sɛ ɛbɛsɛe no adi. Agent a wakɔ anim sen biara mpo behyia tebea horow a ontumi nsi pi wɔ nea ɔde di dwuma no adwene anaa ɔkwan a eye sen biara a ɔbɛfa so ayɛ ho. Sɛnea edi saa adwenem naayɛ yi ho dwuma no yɛ bere a ɛkyerɛkyerɛ mu. Agent a wɔayɛ no yiye nsusuw ho; ɛyɛ kɛse. Adwene mu NhyɛaseSɛ ɔnanmusifo bi gye ne anohyeto ahorow tom sen sɛ obesusuw ho a, ɛma ahotoso denam obu a obu ma nea ɔde di dwuma no tumidi wɔ tebea horow a emu nna hɔ mu no so. Escalation Patterns no bi ne:
Requesting Clarification“Wokaa ‘Next Tuesday.’ So wopɛ sɛ woka September 30th anaa October 7th?” Presenting Options“Mehunuu wimhyɛn mmiɛnsa a ɛne wo gyinapɛn hyia.Emu nea ɛwɔ he na ɛyɛ wo yie?” Nnipa a Wɔde Wɔn Ho Hyehyɛ Mu a WɔsrɛWɔ nnwuma a ɛho hia kɛse anaasɛ emu nna hɔ kɛse no, ɛsɛ sɛ ɔnanmusifo no nya ɔkwan a ɛda adi pefee a ɔbɛfa so akɔ nnipa ho ɔbenfo anaa ɔboafo a ɔboa no mu. Ebia nea ɛbɛkanyan no ne sɛ: “Ɛte sɛ nea saa asɛm yi yɛ soronko, na minni ahotoso wɔ sɛnea mɛkɔ so no ho. So wobɛpɛ sɛ mede frankaa hyɛ eyi so ma onipa nanmusifo bi hwɛ mu?”
Bere a Ɛsɛ sɛ Wɔde Saa Nhwɛso Yi Di Kan Di kan wɔ domɛn a ɔdefo adwene betumi ayɛ nea emu nna hɔ anaasɛ egyina nsɛm a ɛfa ho so kɛse (e.g., abɔde mu kasa nkitahodi, data nsɛmmisa a ɛyɛ den). Fa eyi di dwuma bere biara a agent no de nsɛm a enni mũ bedi dwuma anaasɛ bere a akwan a ɛteɛ pii wɔ hɔ. Asiane a Ɛwɔ OmissionSɛ eyi nni hɔ a, awiei koraa no, ɔnanmusifo no bɛyɛ ahotoso, asiane mu nsusuwii a ɛma nea ɔde di dwuma no twe ne ho fi afoforo ho. Metrics a Ɛma Odi Yiye:
Escalation FrequencyAgent Abisadeɛ a ɛfa Mmoa / Nnwuma nyinaa ho. Akwahosan mu range: 5-15%. Recovery Success RateNnwuma a Wɔawie wɔ Post-Escalation / Total Escalations. Botae a wɔde asi wɔn ani so > 90%.
Nhwɛsode Nea eye sen biara Ma Asiane Titiriw Key Metric Adwene a Wɔde Di Dwuma Nhwɛso Nneyɛe a wontumi nsakra anaasɛ sikasɛm mu nneyɛe User te nka sɛ wɔabɔ no akuturuku >85% Gye a Wogye Tom Autonomy Dial a Wɔde Di Dwuma Nnwuma a asiane dodow a ɛsakrasakra wom Total feature a wogyaee Churn a wɔde si hɔ Ntease a Wɔde Kyerɛkyerɛ Mu Akyi anaa nnwuma a ɛyɛ ne ho Ɔdefo no hu mfomso ahorow “Dɛn ntia?” Tekete Po a Wɔde Di Dwuma Ahotoso Nsɛnkyerɛnne Animdefo anaasɛ nhyehyɛe ahorow a wɔde sika kɛse gu mu Automation bias a wɔde yɛ adwuma Nhwehwɛmu a wɔyɛe Delta Adeyɛ Nhwehwɛmu & Undo Agent nhyehyɛe ahorow nyinaa Ahotoso a wɔhwere daa <5% .Reversion Rate (Nneɛma a Wɔsan Yɛ Escalation Ɔkwan a Wɔfa so Ambiguous user adwene Nsusuwii a ahotoso wom, a ɛde ɔsɛe ba >90% Nkonimdie a Wɔde Yɛ Adwuma
Table 1: Agentic AI UX nhyehyeɛ ho nsɛm tiawa. Kae sɛ wobɛsesa metrics no a egyina wo domain asiane pɔtee ne w’ahiadeɛ so. Designing for Repair na Wɔasiesie Eyi ne sɛnea wobesua sɛnea wobɛpa kyɛw wɔ ɔkwan a etu mpɔn so. Pa kyɛwpa pa gye mfomso no tom, siesie nneɛma a asɛe no, na ɛhyɛ bɔ sɛ obesua biribi afi mu. Mfomso nyɛ nea ebetumi aba; wɔyɛ nneɛma a wontumi nkwati. Sɛnea nhyehyɛe bi a wɔde yɛ adwuma no di yiye bere tenten no nnyina tumi a etumi yɛ pɛ so kɛse, na mmom egyina tumi a etumi san nya ahoɔden wɔ adom so bere a adi nkogu no so. Nhyehyɛe a emu yɛ den a wɔde besiesie nneɛma na wɔasiesie no yɛ ade titiriw, na ɛnyɛ nea wosusuw ho wɔ akyi. Pa kyɛw a Wɔde Tema Yɛ ne Nsiesiei a Ɛda Hɔ Sɛ ɔnanmusifo bi di mfomso a, mfomso nkrasɛm no ne kyɛwpa. Ɛsɛ sɛ wɔde adwene mu pɛpɛɛpɛyɛ na ɛyɛ. Saa bere yi yɛ hokwan titiriw a yɛde bɛda akontaabu adi. Sɛ yɛhwɛ ɔsom nhyehyɛe kwan so a, ɛha na nnwumakuw betumi de ɔsom adwuma a wɔsan nya ho abirabɔ no adi dwuma: adeyɛ a ɛma adetɔfo a onya ɔsom mu huammɔdi, a ɛno akyi no, onya ahoɔden a edi mu na ɔwɔ tema no betumi abɛyɛ nokwaredifo ankasa sen adetɔfo a onhyiaa huammɔdi koraa da. Mfomso a wodi ho dwuma yiye betumi ayɛ adeyɛ a tumi wom kɛse a ɛma wonya ahotoso sen abakɔsɛm tenten a ɛfa kum a mfomso biara nni ho ho. Nea ɛho hia ne sɛ wobebu mfomso no sɛ abusuabɔ a asɛe a ɛsɛ sɛ wosiesie. Eyi hwehwɛ sɛ:
Gye Mfomsoɔ no tomƐsɛ sɛ nkra no ka pefee na ɛyɛ tiawa sɛ wɔadi mfomsoɔ.Nhwɛsoɔ: Mede sika kɔmaa obi foforɔ wɔ ɔkwan a ɛntene so. Ka Nteɛsoɔ a Ɛba Ntɛm araNtɛm ara di nsiesie adeyɛ no akyi.Nhwɛsoɔ: Madane adeyɛ no, na wɔasan de sika no akɔ wo akontaabuo mu. Fa Ɔkwan a Ɛfa Mmoa Foforo Ho MaBere nyinaa fa nkitahodi a emu da hɔ a ɛkɔ nnipa mmoa so ma. Eyi brɛ abasamtu ase na ɛkyerɛ sɛ akontaabu nhyehyɛe bi wɔ hɔ a ɛboro ɔnanmusifo no ankasa so.
Ebia UI a wɔayɛ no yiye a wɔde siesie nneɛma bɛyɛ te sɛ eyi: Yɛdii mfomsoɔ wɔ wo nnansa yi transfer.Mepa wo kyɛw. Mede $250 kɔɔ akontaabu a ɛnteɛ mu.✔ Adeyɛ a Wɔde Teɛ: Wɔadan sika a wɔde kɔmae no, na wɔasan de wo $250 no ama wo.✔ Anamɔn a Edi Hɔ: Wɔahyɛ asɛm a esii no frankaa sɛ wɔmfa nhwɛ mu na amma ansi bio.Whia mmoa foforo? [ Nkitahodi Mmoa ] .
Aban Engine a Wɔbɛkyekye ama Nneɛma Foforo a Ahobammɔ Wɔ Nsusuwii ahorow a yɛaka ho asɛm wɔ atifi hɔ no yɛ nea ɔde di dwuma no anim, nanso wontumi nyɛ adwuma yiye a enni emu mmoa nhyehyɛe a ɛyɛ den. Eyi nyɛ adwumayɛfo akwanside ahorow a wɔbɛbɔ ho asɛm; ɛfa mfaso a ɛwɔ ɔkwan a wɔfa so yɛ adwuma a wɔbɛkyekyere ho. Ahyehyɛde a ɛwɔ nniso nhyehyɛe a ɛho akokwaw betumi de agentic features a ɛyɛ aniberesɛm kɛse akɔma wɔ ahoɔhare ne ahotoso kɛse mu, a wonim sɛ guardrails a ɛho hia no wɔ hɔ de brɛ brand asiane ase. Saa nniso engine yi dan ahobammɔ fi nhwehwɛmu kratasin mu kɔ akansi agyapade. Ɛsɛ sɛ saa engine yi yɛ adwuma sɛ nnisoɔ kuo a ɛwɔ mmara mu, Agentic AI Ethics Council, a ɛwɔ UX, Product, ne Engineering a ɛyɛ adwuma ahodoɔ apam, a mmoa a ɛho hia firi Mmara, Compliance, ne Support hɔ. Wɔ ahyehyɛde nketewa mu no, saa ‘Agyinatukuw’ dwumadi ahorow yi taa hwe ase ma ɛyɛ Product, Engineering, ne Design akannifo abiɛsa biako. Nhwehwɛmu Nhoma a Ɛfa Aban Ho
Mmara/Ahyɛdeɛ Saa kuo yi ne ɔkwan a ɛdi kan a wɔfa so bɔ wɔn ho ban, hwɛ sɛ agent no nneyɛeɛ a ɛbɛtumi ayɛ no bɛtena mmara ne mmara hyeɛ mu. Wɔboa ma wɔkyerɛkyerɛ mmeae a ɛyɛ den a wontumi nkɔ hɔ ma ahofadi a wɔde yɛ ade no mu. ProductThe product manager yɛ agent no atirimpɔw sohwɛfo. Wɔnam ahofadi ho nhyehyɛe a ɛwɔ hɔ a ɛkyerɛw nea ɔnanmusifo no yɛ ne nea wɔmma no kwan sɛ ɔnyɛ so na ɛkyerɛkyerɛ ne dwumadi ahye so na wɔhwɛ so. Wɔn na wɔwɔ Agent Risk Register no. UX NhwehwɛmuSaa kuw yi yɛ nne a ɛkyerɛ nea ɔde di dwuma no ahotoso ne ne dadwen. Wɔn na wɔhwɛ nhyehyɛe a ɛsan ba bio a wɔde yɛ ahotoso calibration adesua, subammɔne sɔhwɛ ahorow a wɔayɛ no sɛnea ɛte, ne nsɛmbisa a ɛfa su ho de te user’s evolving mental model of the agent ase. EngineeringSaa kuw yi kyekye mfiridwuma mu nnyinaso a ɛma ahotoso. Ɛsɛ sɛ wɔyɛ nhyehyɛe no ma logging a ɛyɛ den, undo functionality a wɔde klik pɛnkoro, ne hooks a ehia na ama wɔanya ntease ahorow a emu da hɔ na wotumi kyerɛkyerɛ mu. MmoaSaa akuo yi wɔ anim wɔ huammɔdi mu. Ɛsɛ sɛ wɔtete wɔn na wɔsiesie wɔn ma wodi nsɛm a esisi a efi ananmusifo mfomso mu ba no ho dwuma, na ɛsɛ sɛ wonya nsɛm a wɔka kyerɛ tẽẽ kɔma Abrabɔ Pa Bagua no de bɔ wiase ankasa mu huammɔdi nhyehyɛe ahorow ho amanneɛ.
Ɛsɛ sɛ saa nniso nhyehyɛe yi kura ankrataa a nkwa wom a wɔahyehyɛ, a Agent Risk Register a ɛde nnam kyerɛ akwan a ebetumi adi huammɔ, Action Audit Logs a wɔtaa hwɛ mu, ne Autonomy Policy Documentation a ɛyɛ mmara kwan so ka ho. Baabi a Wobefi Ase: Ɔkwan a Wɔfa so Fa Nneɛma a Wɔyɛ no Akannifo a Wɔyɛ no Nkakrankakra Wɔ nneɛma so ahwɛfo ne mpanyimfo fam no, agentic AI a wɔde bɛka abom no betumi ate nka sɛ ɛyɛ adwuma kɛse. Nea ɛho hia ne sɛ yɛbɛbɛn no ɛnyɛ sɛ nea wɔde bɛto gua biako, na mmom sɛ akwantu a ɛkɔ so nkakrankakra a wɔde kyekye mfiridwuma mu tumi ne ahotoso a wɔde di dwuma no nyinaa wɔ bere koro mu. Saa kwankyerɛ yi ma w’ahyehyɛde no tumi sua na ɛsakra, hwɛ hu sɛ wɔde anammɔn biara asi fapem a ɛyɛ den so. Ɔfã 1: Fapem Ahobammɔ (Fa nyansahyɛ & Ho nyansahyɛ) . Botae a edi kan ne sɛ wɔbɛkyekye ahotoso nnyinaso a wɔremfa asiane akɛse a wɔde wɔn ho nto wɔn ho so. Wɔ saa fã yi mu no, agent’s tumi no yɛ nhwehwɛmu ne nyansahyɛ nkutoo.
Fa Intent Preview a ɛyɛ den te sɛ ɔbotan di dwuma: Eyi ne wo nkitahodi titiriw nhwɛso. Ma wɔn a wɔde di dwuma no ho tɔ wɔn wɔ adwene a ɛne sɛ ɔnanmusifo no reyɛ nhyehyɛe ahorow no ho, bere a woma nea ɔde di dwuma no di dwuma koraa wɔ kum no so. Build the Action Audit & Undo infrastructure: Sɛ mpo agent no nyɛɛ ade wɔ ne ho mu de besi nnɛ a, si technical scaffolding no ma logging ne reversal. Eyi siesie wo nhyehyɛe no ma daakye na ɛma nea ɔde di dwuma no nya ahotoso sɛ ahobammɔ afiri bi wɔ hɔ.
Ɔfã 2: Calibrated Autonomy (Adeyɛ a Wɔde Si so dua) . Sɛ wɔn a wɔde di dwuma no ho tɔ wɔn wɔ agent’s proposals ho wie a, wubetumi afi ase de ahofadi a asiane nnim aba. Saa ɔfa yi fa sɛnea wɔbɛkyerɛkyerɛ wɔn a wɔde di dwuma no sɛnea ɔnanmusifo no susuw nneɛma ho na wɔama wɔn ankasa de wɔn ahoɔhare asi hɔ.
Fa Autonomy Dial a nhyehyɛe a anohyeto wom no ba: Fi ase denam ma a wobɛma wɔn a wɔde di dwuma no kwan ma wɔama agent no tumi sɛ ɔnyɛ Ade a Wɔde Si so dua no so. Deploy the Explainable Rationale: Wɔ adeyɛ biara a ɔnanmusifo no siesie ne ho no, fa nkyerɛkyerɛmu a emu da hɔ ma. Eyi yi agent’s logic no fi ahintasɛm mu na ɛhyɛ mu den sɛ egyina nea ɔde di dwuma no ankasa pɛ so na ɛyɛ adwuma.
Ɔfã 3: Nnanmusini a Wɔde Di Dwuma (Ayɛ Adwuma a Wɔwɔ Wɔn Ho) . Eyi ne anammɔn a etwa to, a woyɛ bere a woanya data a emu da hɔ a efi afã horow a atwam no mu a ɛkyerɛ sɛ wɔn a wɔde di dwuma no wɔ nhyehyɛe no mu ahotoso akyi nkutoo.
Ma Act Autonomously nyɛ adwuma ma nnwuma pɔtee bi a wɔadi kan apene so: Fa data a efi Phase 2 (e.g., Proceed rates a ɛkorɔn, Undo rates a ɛba fam) di dwuma de kyerɛ nnwuma a asiane nnim a wobetumi de ayɛ adwuma koraa. Monitor and Iterate: Autonomous features a wɔde bɛba no nyɛ awiei, na mmom ɛyɛ mfiase a ɛkɔ so kyinhyia a wɔde hwɛ adwumayɛ so, boaboa wɔn a wɔde di dwuma no nsɛm ano, na wɔsiesie agent no kɛse ne ne nneyɛe a egyina wiase ankasa data so.
Design Sɛ Ahobammɔ Lever a Etwa To Agentic AI a ɛbae no gyina hɔ ma ɔhye foforo wɔ nnipa ne kɔmputa nkitahodi mu. Ɛhyɛ daakye a mfiridwuma betumi de nsiyɛ atew yɛn adesoa so na ama yɛn asetra ayɛ mmerɛw ho bɔ. Nanso tumi yi de asɛyɛde a emu dɔ ba. Ahofadi yɛ nea efi mfiridwuma nhyehyɛe bi mu ba, nanso ahotoso yɛ nea efi nhyehyɛe bi mu ba. Yɛn asɛnnennen ne sɛ yɛbɛhwɛ sɛ osuahu a ɔde di dwuma no nyɛ mfiridwuma mu tumi a wɔapirapira na mmom nea obenya so mfaso titiriw. Sɛ́ UX adwumayɛfo, nneɛma so ahwɛfo, ne akannifo no, yɛn dwumadi ne sɛ yɛbɛyɛ saa ahotoso no sohwɛfo. Ɛnam nhyehyeɛ a ɛda adi pefee a yɛde bedi dwuma ama ahyɛnsodeɛ ne pene, akwan a wɔasusu ho a yɛbɛfa so asiesie, ne nnisoɔ nhyehyɛeɛ a ɛyɛ den a yɛbɛkyekyere so no, yɛbɔ ahobanbɔ nhama a ɛho hia a ɛma agentic AI tumi yɛ adwuma. Ɛnyɛ sɛ yɛreyɛ interfaces kɛkɛ; yɛreyɛ abusuabɔ ahorow ho nhyehyɛe. AI’s mfaso ne gye a wogye tom daakye gyina yɛn tumi a yɛde nyansa, nhumu a ɛkɔ akyiri, ne obu a emu dɔ a yɛwɔ ma nea ɔde di dwuma no tumi a etwa to no bɛyɛ nhyehyɛe ahorow a ɛyɛ den yi so.
