Business operations

AI adoption by role

Business operations is where AI (artificial intelligence) meets customers' real problems: the contact center, the back office, branch and servicing teams. These staff use AI-enabled tools in live work, often under time pressure. They are also the people most likely to notice when an answer is wrong. This track covers two seats: the operator who uses the tool, and the operational owner who decides how the unit uses it.

The rule for this seat: AI output is decision support inside your documented authority. You check the cited source and the case facts, you follow the fallback when the tool abstains or fails, and you report what went wrong. Counts and thresholds on this page are illustrative.

1. At a glance

Item	For business operations
Mandate	Serve customers correctly using approved procedures; use AI-enabled tools only within your authority; report errors and near misses
Primary decision	Relay, act or escalate: is this answer verified, and is it mine to give?
First safe win	Use the internal policy assistant for a policy question, open the cited section, and confirm it before you answer
Biggest failure mode	Automation bias under pressure: relaying a fluent answer without opening the source, or pasting customer details into an unapproved tool to get a faster answer

Line: 1st line · Family: Adopt · Variants: contact center, back office, branch and servicing; operator and operational owner

Variant	What differs
Contact center	Live calls and chats. Seconds count, so verification must be fast. The policy assistant is the main AI tool here
Back office	Casework without a customer waiting: summaries of case history, drafts of internal notes. More time to verify; more volume, so sampling matters
Branch and servicing	Face-to-face service. The customer can see the screen; disclosures and identity checks must happen in the right order
Operator	Uses the tool. Owns each relay, action and report. Cannot change the procedure
Operational owner	Owns the unit's procedure, training, fallback, staffing and the decision to pause use. Is the business owner on the intake card for tools like the policy assistant

2. Your journey: first steps to fully adopted

The journey for business operations starts with safe habits at the desk and ends with a unit where AI supports every answer and people still make every decision.

Start here: your first steps

Pass the acceptable-use call scenarios before anyone gets access.
Switch on near-miss reporting and thank the first people who use it.
Ask the policy assistant one policy question, and open the cited section before you answer.
Publish the unit procedure and drill the fallback path with the tool switched off.

Fully adopted: how the work runs

Moment in the work	Before AI	Fully adopted (Stage 4–5)	Decision that stays human
A customer asks	The operator searches the procedure manual or asks a colleague	The operator asks the policy assistant in general terms; it answers with a citation or abstains	What to ask, with no customer details in the prompt
The answer	Answers vary by person and by shift	The operator opens the cited section and relays the answer in their own words	Relay, act or escalate
Account facts	Read from the servicing system	Still read from the servicing system: the due date and cutoff come from payment-schedule-service, never from the assistant	Every account action, identity check and disclosure
Case notes	Typed by hand after the contact	AI drafts the internal case summary; the operator reviews and signs it	Signing each summary
Improvement	Wrong answers repeat for months	Near misses and corrections feed the weekly review; fixes reach the policy index or the tool	Pausing the tool in the unit, by the operational owner

What changes for the organization

Before	After
Answers depend on who picks up	The same cited answer on every shift
Errors found through complaints	Errors found early, through near-miss reports
Speed targets punish checking	Verification time is built into handle-time targets

In any organization

Small or unregulated: the support lead or the owner covers this seat; operator and operational owner may be the same person.
What scales: formal attestation and the near-miss register shrink (a shared form works). No customer data in unapproved tools, a person deciding every customer action, a fallback when the tool is off and the ability to switch it off never do.

3. Your decisions

Decision relationships on this site: owner or approver, independent validator or challenger, consulted control partner, evidence recipient, notified.

Decision	You own / advise / challenge	Inputs	Escalation	Cannot be delegated
Relay an answer to a customer	Operator owns	Assistant answer, the cited section, case facts	Team lead	Yes
Take an account action	Operator owns, within existing authority; the assistant never acts	Procedure, case facts, required verification	Team lead	Yes
Escalate or use the fallback	Operator owns	Abstention, doubt, a mismatch	Team lead, supervisor	Yes
Report a near miss	Operator owns	What happened, without customer data in free text	Operational owner	Yes
Unit procedure for AI use	Operational owner owns; compliance consulted	Policy, evaluation results, training	Business owner, compliance	Yes
Pause the tool in the unit	Operational owner owns; production support may disable first in an incident	Near-miss trend, complaints, monitoring	AI risk, production support	Yes
Policy content	Challenge; the policy office owns it	Wrong or stale answer traced to a document	Policy office	Yes

4. Two kinds of adoption

Kind	What it means for business operations	Example
How you work with AI	Use approved tools for questions, summaries and internal drafts. Verify before you relay. Never paste customer data where it is not approved	Ask the policy assistant about the cutoff rule on a holiday; open the cited section; then answer
What you operate that contains AI	The operational owner runs an AI-enabled service: procedure, training, fallback, monitoring of corrections, and the call to pause	The unit's procedure for the policy assistant, its fallback when it is off, and the weekly near-miss review

5. Your ladder

The ladder here is the unit's, set by the operational owner. An operator does not climb alone.

Stage	AI use	Guardrail	Artifact or evidence	Decision relationship	Step-down trigger
0 · Foundations	None. Training on acceptable use with call scenarios	Acceptable use; data classification; escalation path published	Attestation; scenario check passed	Operator attests; operational owner tracks	n/a
1 · Personal	Approved assistant for general, non-customer questions	No customer data in prompts	Near-miss reporting switched on	Operator; operational owner notified of reports	Any unreported breach
2 · Team context	Policy assistant from the approved, versioned policy index, with citations	Procedure: verify the cited section before relaying a consequential answer	Unit procedure; verification spot checks	Operational owner owns the procedure; compliance consulted	Policy index out of date; spot checks fail
3 · Delegated	Back office: AI drafts internal case summaries an operator reviews	Evaluation set for summaries; operator signs each one	Signed summaries; sample review results	Operator owns each summary; team lead samples	Sample error rate above threshold
4 · Agentic	Not used for customer actions. Any step toward account actions is a new intake, likely High	Intake and High-tier controls before any pilot	Intake card	Business owner; control partners assign the tier	n/a: no customer action without a new intake
5 · Scale	Shared procedures, near-miss taxonomy and training across units	Periodic review; material-change process	Shared procedure entries	Operational owners; enablement lead supports	Material change in tool, policy or population

Stays human at Stage 5: the decision to relay, every account action, identity verification, required disclosures, and complaint handling. The assistant never takes an action.

6. Priority scenarios

Scenario	Value	Risk tier	Stage	First step	Do not use AI when…
Policy question during a call	Faster, consistent answers	Medium (staff verify before relaying)	2	Ask; open the cited section; then answer	The answer would be relayed without the required source check: stop, use the fallback, and report the control breach. A design that sends AI output directly to customers is a new High-tier intake
Internal case summary	Less reading time	Medium	3	Operator reviews and signs each summary	The summary will be sent to the customer
Customer letter draft	Faster correspondence	High (reaches a customer)	Intake first	Use approved templates; any AI draft goes through High-tier controls	No approved High-tier route exists
Explaining a fee or eligibility	Fewer escalations	Medium as support; the decision is the procedure's	2	Verify the cited section and the case facts	The question is about waiving, approving or denying: follow the procedure
Hardship or adverse decision	None from AI	Prohibited if automated	Never	Follow the existing procedure	Always: a fully automated adverse decision is prohibited
Near-miss reporting	Improves the tool for everyone	Low	1 onward	Use the correction control or the report form	Never skip it to save time

7. The due-date ask from your seat

The contact center sees the due-date problem first, as calls. Full narrative below; the delivery side is on the worked example. A representative here means a person on the phone, not an AI agent.

Fact the representative needs	What it means on a call
Due date	A date-only value from the contract, held by payment-schedule-service. It is the same date for every customer in every zone. "Nov 16" is "Nov 16"
Cutoff	An instant: 5:00 PM Eastern on the due date. A customer in Los Angeles hears "2:00 PM Pacific time"
Fallback label	"Based on your state of residence" means the profile has no zone. Offer to confirm the customer's zone; do not guess
Where to read it	The representative reads the authoritative due date and cutoff already exposed in the servicing system. The contact-center desktop is outside this release

When the channels disagree. A customer says the app showed Nov 15 but the statement says Nov 16.

Step	Representative does	Why
1 · Read the source	Reads the due date from the field the procedure names (sourced from payment-schedule-service)	That is the contractual date
2 · Answer	"Your due date is Nov 16. The cutoff is 2:00 PM Pacific time on Nov 16."	Due date and cutoff stated separately, cutoff in the customer's zone
3 · Do not fix it yourself	Does not change any date or promise a fee change outside the procedure	No authority to change records
4 · Tag and report	Tags the contact "due date confusion"; files a discrepancy report: channel, date seen, date shown in the source, time of the view	Feeds the monitoring signal owned by the contact-center lead; the report goes to the team as a defect
5 · Remediate by procedure	If a late fee followed from the wrong date, follows the existing remediation procedure	Customer harm is handled by the procedure, not by AI

If the policy assistant is asked "when is the cutoff?", it answers the rule from policy with a citation. It cannot see the account. The representative reads the account's dates from the servicing system, never from the assistant.

The policy assistant, concretely

The companion policy assistant is Medium tier because staff verify the cited source before relaying. Here is what that means at the desk.

Moment	Operator does	Never
Ask	Types the policy question in general terms: "cutoff rule when the due date is a holiday"	Types names, account numbers or other customer details
Answer arrives	Opens the cited document, version and section; reads the passage	Relays a fee, cutoff or eligibility answer without opening the passage
Passage matches	Relays the answer in their own words	Reads AI text aloud as if it were the policy
Passage does not match	Uses the fallback; files a correction	Picks whichever sounds right
Abstention ("I can't answer that from approved policy. Ask your supervisor.")	Follows the fallback path: procedure manual, then supervisor	Rephrases the question until it gives an answer
Assistant is off	Uses the procedure manual and the supervisor line as before	Switches to an unapproved public chatbot
Something feels wrong	Reports a near miss, even if no customer was affected	Stays quiet to keep handle time down

8. Hand-offs and separation of duties

Direction	What	With whom
You give →	Corrections, near misses, discrepancy reports	Operational owner, AI risk intake, the delivery team
You give →	Contact tags and complaint themes	Product owner, contact-center lead
You give →	Process walk notes and exception lists	Business analyst
You receive ←	Procedures, training, the fallback path	Operational owner, enablement lead
You receive ←	Policy documents	Policy office
You receive ←	Answer card design and abstention wording	Experience designer
You challenge	Any answer whose citation does not support it; any target that pushes speed over verification	Operational owner
You are notified of	Tool pauses, policy changes, procedure changes	Operational owner, production support

The operator relays and acts; the operational owner sets the procedure; the policy office owns the policy; the AI team owns the tool. None of them approves their own change.

9. Evidence and records

Evidence produced or consumed	System of record	Retention owner
Acceptable-use attestation (produced)	Learning system	Enablement lead
Near misses and corrections (produced)	Near-miss register in the monitoring system	AI risk office
Discrepancy reports and contact tags (produced)	Case and complaint system	Operational owner
Signed case summaries (produced)	Case system	Operational owner
Assistant logs: question, cited sections, answer, abstention (consumed in reviews)	Assistant log store, customer identifiers masked	Platform team, under the records schedule
Unit procedure and training records (produced by the owner)	Procedure library; learning system	Operational owner

10. Skills and learning path

Level	You can…	Practice
Foundation	Pass acceptable-use call scenarios; tell a cited answer from an uncited one; name what never goes in a prompt	Scenario: a customer reads out an account number while you type a question
Practitioner	Verify a cited section in under a minute; use the fallback; file a useful near miss	Practice rounds on the Lab L6 assistant with planted wrong citations
Advanced (operational owner)	Write the unit procedure; read the near-miss trend; decide when to pause	Run the weekly review with AI risk; use the adoption scorecard

11. Measures and incentives

Metric	Definition	Anti-gaming note
Verification rate	Share of consequential answers where the cited passage was opened before relay	Pair with spot checks; opening is not reading
Near-miss reports	Reports per hundred staff per month	Rising reports are good news early on. Never penalize reporters
Correct abstention follow-through	Share of abstentions that went to the fallback path	Rephrasing to force an answer counts as a miss
Due-date confusion contacts	Contacts tagged "due date confusion"	Should fall after release; check tagging is consistent
Handle time	Time per contact	Never a target on its own: it rewards skipping verification

12. Cost, capacity and workforce

Item	Note for the operational owner
Verification time	Build the check into handle-time expectations. If targets do not allow it, staff will skip it
Training time	Scenario training before access, refreshers after policy or tool changes
Job concerns	Say clearly what the tool does and does not change in the role. Plan it with the enablement lead and HR (human resources)

13. Anti-patterns and stop conditions

Anti-pattern or trigger	Why it hurts	What you do
Relaying without opening the source	Turns a Medium tool into an unverified High one	Stop; verify; retrain if it is a pattern
Customer data in an unapproved tool	Data leaves the organization	Stop; report the near miss; use approved channels only
Rephrasing past an abstention	Forces an unsupported answer	Use the fallback path
Speed targets that punish checking	Rewards automation bias	Operational owner changes the target
Spike in corrections or complaints	The tool or the policy index may be wrong	Operational owner pauses use in the unit and notifies AI risk
Policy, model or population change	Old training and evaluation may not hold	Refresh training; wait for re-evaluation before resuming

14. Action kit and 30/60/90

## Operator · policy assistant desk card · v1
1. Ask in general terms. No names, account numbers or case details.
2. Open the cited document, version and section. Read the passage.
3. Matches? Answer in your own words. Fees, cutoffs, eligibility: always check.
4. No match, or "I can't answer that"? Procedure manual, then supervisor.
5. Account dates come from the servicing system, never from the assistant.
6. Due date is a date: the same everywhere. Cutoff is a time: say it in the customer's zone.
7. Channels disagree? Read the source field, tag "due date confusion", file a discrepancy.
8. Anything odd: report a near miss. No blame for reporting.

Questions to ask your enablement lead:

Which AI tools are approved for my unit, and which are explicitly not?
How long should verification take, and is that time inside my handle-time target?
Where do near misses go, and what happened to the last ones we reported?

Window	Goals	Done when
First 30 days	Stage 0: scenario training; near-miss reporting on	All staff attested; first reports filed
Days 31 to 60	Stage 2: policy assistant with the desk card and the unit procedure	Verification spot checks pass; fallback drilled with the tool off
Days 61 to 90	Operational owner runs weekly near-miss reviews; feeds fixes back	Two fixes traced from reports to the policy index or the tool