Adapt AI adoption to the work people do and the consequences of getting it wrong. Keep a shared foundation for approved tools, data handling and support, then test specific tasks with the people who perform and receive the work. A job title or industry label is a starting point for that conversation, not an acceptance test.
For an adoption lead coordinating several departments, the useful question is what must change when a practice moves. A finance analyst, store supervisor and recruiter may all ask for a summary. They need different source material, different checks and different authority to act on the answer.
This guide helps you make those differences explicit. Use the AI adoption strategy guide for the wider program and the tools and workflows guide for choosing and supporting the technology.
Start with the task inside the role
“AI for finance” is too broad to evaluate. Preparing a first draft of an invoice discrepancy note is a more useful unit of work. Someone can inspect the documents, identify an error and decide whether the note is ready to send.
Suppose a retailer's finance analyst compares a supplier invoice with the purchase order and the warehouse's delivery record. The purchase order says what the retailer agreed to buy; the delivery record says what arrived. An assistant might organize the differences and draft a query. The analyst still checks quantities, prices and document references before asking the buyer to resolve the mismatch. Approval to draft that query does not grant authority to approve payment.
The same exercise exposes different needs elsewhere. These are possible trial tasks to investigate with each department, not assumptions about what every team should automate.
| Role and possible trial | Output to inspect | Important boundary |
|---|---|---|
Finance analyst comparing invoice records | A discrepancy note linked to the source documents | Checking the note and authorizing payment are separate decisions |
Recruiter preparing an interview | Questions grounded in the approved role requirements | Drafting questions does not authorize ranking or rejecting candidates |
Sales representative preparing an account meeting | A brief with sourced customer facts and unresolved questions | Do not turn an inference into a customer commitment |
Support agent drafting a reply | An answer checked against current support guidance | Exceptions go to the person authorized to decide them |
Engineer investigating a software defect | A proposed change with a reproducible test | A plausible explanation is not evidence that the fix works |
Store supervisor preparing a shift briefing | Instructions matching the current approved store procedure | A summary must not invent a new policy |
Evidence from one task should travel cautiously. In a 2023 study described by Harvard Business School, 758 BCG consultants took part in experiments with GPT-4. AI improved performance on selected product-development tasks, but reduced correct answers on a different business problem designed to be outside its capabilities. The result concerns those tasks and that model, not the performance of today's tools. It shows why even one occupation cannot be treated as a single test case.
Before a trial, name the source material, accepted output, reviewer and action boundary. If the team cannot agree on those, observe a recent piece of work together before choosing a demonstration.
Ask what the industry changes
An industry can change the data available, the people affected, the operating conditions and the decisions that require specialist authority. Those conditions matter more to a local trial than a sector's position in an adoption ranking.
A Census Bureau working paper illustrates the difference between broad uptake and internal spread.
57%
of firms adopting AI across business functions used it in three or fewer functions
The April 2026 working paper analyzed the U.S. Business Trends and Outlook Survey's AI supplement, with a November 2025 to January 2026 reference period. Firm respondents reported use; this is a measure of breadth across functions, not the share of employees using AI or proof of business value.
The paper also finds substantial variation by firm size and sector. Treat such evidence as context for asking better questions. It cannot tell you whether your store's returns summary or your finance team's discrepancy note will be useful.
Use the following starting points to identify what needs local investigation. Confirm applicable requirements with the people responsible for them before using sensitive records or changing a consequential decision.
Banking and insurance
Separate employee assistance from decisions affecting a customer. For a service representative's draft explanation of a product, identify the approved wording, source version and reviewer. A trial of that draft does not establish permission to decide eligibility, price a policy or make a credit decision. Ask which customer records the task actually needs and who can approve their use.
Healthcare
Distinguish administrative work, documentation and clinical decisions. A hospital administrator drafting a staff-training invitation from an approved schedule faces a different task from a clinician interpreting patient information. Define that boundary first, then involve the appropriate clinical, privacy and operational owners for any proposed expansion.
Manufacturing and industrial work
Identify where a written output could influence a physical action. Summarizing a maintenance history is different from issuing instructions to operate or repair equipment. Ask the responsible engineering or safety owner what must remain authoritative, how workers check the source and how a shift handoff preserves unresolved questions.
Retail and hospitality
Examine access during the shift. Does a store associate have an approved device, time to use it and a way to ask for help while serving a customer? Test a task such as preparing a returns-policy explanation against the current store procedure. Include an exception requiring supervisor review, rather than demonstrating only the straightforward case.
Law, accounting and consulting
Clarify the distinction between preparation and professional sign-off. For a client briefing, identify which statements need supporting authority, what client material may be used and who accepts the final advice. Examine what junior staff still need to understand and practise when a tool produces the first draft. Faster preparation does not settle those responsibilities.
Public administration
Consider the person who must understand or challenge the result. A council employee drafting a response about a published service can check it against the current service information. A decision about a resident's entitlement is a different scope. Ask the service owner how accessibility, explanation, records and escalation will work for the proposed use.
Construction, logistics and field services
Test the handoff between office and site. A dispatcher may have a desktop and complete records while a driver or site supervisor has intermittent access. For a delivery-exception summary, confirm the location, original report and person responsible for the next action. Keep existing safety and urgent-incident procedures authoritative.
Education, nonprofits and research
Separate staff administration from judgments about learners, beneficiaries or scientific findings. Preparing a facilities booking summary from approved requests can be a bounded trial. Assessing a learner or interpreting research evidence needs its own standards and responsible reviewers. Ask whose information is involved and whether the output will be mistaken for an authoritative judgment.
Let the work shape the practice
A shared introduction can explain the organization's permitted tools and basic checking habits. The next practice session should use work the participant recognizes, with an output they can evaluate.
For the retail example, suppose a supervisor and store associate practise explaining a return under an approved policy. They use a sample purchase record and a plain boxed item. The associate checks the draft explanation against the policy, notices a missing receipt and identifies when to ask the supervisor. The exercise tests recognition of an exception as well as fluent wording.
Support can look different even among organizations using similar technology. In a June 2026 Reddit discussion, one Reddit user described a training package covering acceptable use, policy and introductory Copilot material before granting access. They also said the material needed updating. In the same discussion, another Reddit user described a different arrangement.
None at all. We have an in house LLM and people have picked it up quite easily due to how we integrated it into our in house tools.
The commenter was answering a question about formal employee training. In a later reply they described substantial custom infrastructure and cautioned another commenter against copying that route. Employer, company size and outcomes were not independently verified.
These accounts do not establish which approach works better. They make the local question concrete: what does the employee need to learn because of this particular task and implementation? An embedded interface may remove some tool instruction while leaving source checking, exceptions and escalation essential.
Check the conditions for practice with the people involved:
- Access. Can they reach the approved tool and source material with their normal account and device?
- Time. Can they practise without abandoning service, shift or deadline responsibilities?
- Judgment. Do they know what an acceptable output looks like and when to seek a specialist?
- Support. Is help available when the task occurs, including outside the central team's working hours?
The people and change guide covers the wider support system. Here, use those questions to adapt one practice to the role and setting.
Compare outcomes fairly across roles
Do not make identical weekly usage a proxy for equal progress. A finance analyst may need a tool during a monthly close; a support agent may encounter a suitable task throughout the day. Count opportunities to use an approved practice, and inspect whether the resulting work is acceptable.
Gallup's February 2026 survey of 23,717 U.S. employees found differences in frequent use across role levels, while also identifying associations with workflow fit and manager support. Those are self-reported associations, not evidence that a job title fixes someone's willingness or ability to learn.
For each role, agree a relevant outcome and keep the important burden visible. The finance analyst's note should identify the right discrepancy without creating extra correction work. The store associate's explanation should match policy and route the exception correctly. Neither outcome is captured by counting prompts alone.
Use the measurement guide for the distinction between task improvement and business value. When comparing teams, record differences in task frequency, access, review requirements and workload before interpreting a lower use rate as resistance.
Record what changes before transferring a practice
A successful trial is useful to another department when its conditions are visible. Share the example, the checks and what failed, then ask the receiving team to identify the differences.
| Adaptation question | What to record before reuse |
|---|---|
What is the actual work? | Role, inputs, accepted output and recipient |
What changes in this setting? | Data, device access, language, operating hours and consequences of error |
What remains common? | Approved tools, data rules, reporting route and shared support commitments |
Who can judge the result? | Named local reviewer, specialist questions and authority to approve use |
What must the local trial include? | Ordinary examples, important exceptions and a comparison with the current process |
What decision follows? | Continue, adapt or stop, with evidence and a review trigger |
Start with one practice that has a willing local owner and a meaningful way to judge the output. Resolve missing access or authority before expanding the trial. When the work changes enough, treat it as a new test rather than assuming the old result still applies.
Questions about AI adoption by role and industry
Which industries have adopted AI the most?
There is no single ranking that answers every comparison. Results depend on country, firm size, survey date and what counts as AI use. The Census working paper finds higher use among large firms and knowledge-intensive sectors in its U.S. data. Use that as context, then evaluate a specific task in your organization rather than treating a sector's adoption rate as a target or evidence of value.
Does every department need a different AI tool?
Not necessarily. Departments may share an approved tool while using different source material, examples and acceptance checks. Consider a separate tool when the task's capabilities, access restrictions or integration needs justify it. First define the work and reviewer, then compare options under the same conditions using the tools and workflows guide.
What should role-specific AI training include?
Role-specific training should help a person complete and check a recognizable piece of work with permitted inputs. Include a normal example, an important exception, the source used for verification and a route to help. Adapt device access and practice time to the working setting. The retail practice example shows why recognizing a missing receipt matters alongside producing a clear explanation.
Does low AI use mean frontline employees are resisting change?
Low use alone cannot establish resistance. Employees may lack access, have few suitable tasks, be following a restriction or find that checking takes too much effort. Ask about the work and observe an approved attempt before deciding what support is needed. Compare opportunities and outcomes, rather than requiring every role to reach the same prompt count.
Can an AI practice transfer from one industry to another?
A practice can provide a starting point, but its result does not transfer automatically. Compare the input data, accepted output, review authority, operating conditions and consequences of error in the receiving setting. Keep applicable local requirements with their owners and run a bounded test. Use the adaptation record to decide what can be reused and what needs new evidence.
