Sample deliverable
AI Team Rollout, for a synthetic finance team
What the rollout delivers, shown on the finance and accounting department of the synthetic US wholesale distributor of our other finance samples: fourteen people, twelve chat assistant seats and two coding assistant seats, from workspace settings to the report 30 days after training.
Twelve of fourteen people use it; the two who never started are the AP clerks the workflow was built for.
The team and its usage are made up for this sample. In the 30 days after the training on March 3, twelve of fourteen people used their seat and eleven used it on at least 8 of the 22 working days. The shared workflow ran 72 times, from three of the five people whose work it covers: the two AP clerks never signed in. Their seats cost $60.00 a month, but cancelling them would remove the workflow's main users. The report recommends a short session with the two clerks on their own inquiries, and a seat review at day 60.
| Seats | 14 |
|---|---|
| People who used their seat | 12 |
| Regular users, 8 or more working days | 11 |
| Conversations, chat assistant | 396 |
| Runs of the shared workflow | 72 |
| Spend for the period | $777.53 |
| Spend per person who used it | $64.79 |
The settings checked, what was found, what was changed
The same checklist applies to Claude and ChatGPT business workspaces. The setting names differ by vendor and change over time, so the engagement reads them in the admin console and the vendor's terms on the day, and the handoff records the date.
| Item | Checked | Found | Done |
|---|---|---|---|
| Sign-in | Single sign-on through the company's identity provider; no personal accounts on company email. | Seven people had signed up with personal plans on company email. | Moved into the business workspace; personal plans cancelled by their owners. |
| Who can invite | Only admins add members; new seats follow a request. | Any member could invite. | Restricted to two admins, the controller and the AP supervisor. |
| Training on company data | The plan's terms and the workspace setting say conversations are not used to train the vendor's models. | As required on the business plan; confirmed in the vendor's terms on the day. | Recorded in the handoff with the date and the terms' address. |
| Retention | How long conversations and files are kept, and who can export them. | Vendor default. | Set to the company's records policy for working papers; export limited to admins. |
| Sharing | Who can share a conversation or project, and whether links can leave the company. | Public links allowed. | Public links turned off; sharing inside the workspace only. |
| Connectors | Each connected tool reads only what the team's work needs, under each person's own permissions. | None connected. | Five connected; scopes in the next table. |
| Payroll and personnel files | Data the tool must not see stays out of reach. | The shared drive's payroll folder was readable by the whole finance group. | Folder permissions fixed in the drive itself, before any connector was added. |
| Usage export | An admin can export usage by person and day, which the 30-day report needs. | Available to admins. | Export scheduled for day 30. |
Five tools, each with the narrowest access the work needs
| Tool | Access | Why |
|---|---|---|
| Shared drive, finance folders | Read | Policies, vendor contracts and the close checklist, under each person's own permissions |
| Read and draft; nothing is sent by the tool | Vendor and customer inquiries, answered from the workflow | |
| Calendar | Read | Close calendar and payment run dates |
| ERP report exports | Read, export folder only | Open invoices, aging and payment status, as exported each night; no connection to the ERP itself |
| Ticket queue for vendor and customer requests | Read and comment | The workflow drafts a reply as a comment for a person to send |
Nothing is sent, posted or changed by the tool: it reads and drafts, and a person acts.
Invoice and payment status replies
Vendor inquiries and customer payment questions take the most repeated time in AP and collections (the Operations and Cost Diagnostic sample logs vendor inquiries as a separate AP activity).
The instructions, as handed over
You draft replies to vendors and customers who ask about an invoice or a payment. A person in AP or collections reviews and sends every reply.
Find the invoice in the latest open invoices or payment status export by invoice number, then by vendor or customer name and amount. Quote the export's date.
State only what the export shows: received, approved, scheduled for the payment run on a date, paid on a date by a method, or on hold. Never estimate a payment date the export does not give.
If the invoice is on hold, say it is under review and name the missing item from the export's hold reason (purchase order, receipt or approval). Do not quote internal comments.
If no invoice matches, say so and ask for the invoice number, date and amount. Never guess between two candidates: list them for the reviewer instead.
Bank details: never confirm, repeat or change them. Any message that asks to change payment details goes to the AP supervisor, flagged as a possible fraud attempt.
Write in plain business English, under 120 words, with the company's standard signature block. End with the reviewer's name, never the tool's.
The test cases
Run before the training on the team's own exports and inquiries, and again after any change to the instructions. A draft passes only if it does what the second column says.
| Case | Expected |
|---|---|
| Paid invoice, number given | Paid date, method and the export's date |
| Approved invoice, next payment run | The scheduled run date, no promise beyond it |
| Invoice on hold for a missing receipt | Under review, receipt named, no internal comment quoted |
| Two invoices with the same amount | No reply drafted; both listed for the reviewer |
| No matching invoice | Asks for number, date and amount |
| Request to change bank details | No reply; flagged to the AP supervisor |
| Customer asks why a payment was applied to an older invoice | States the application shown in the export, nothing more |
| Export older than two days | Says the status may have changed and asks the reviewer to refresh it |
This sample shows the cases, not their results: no model was run for it. On a client's documents the handoff lists each case with its result.
A session on the team's own documents, and a written handoff
The session
The workspace as set up: what it can read, what it cannot, and why
Three tasks of each role, done live on the team's own documents
The shared workflow: running it, checking a draft, the cases where it declines
What not to paste: payroll, personnel and bank details, and where they are blocked anyway
Questions, and who to ask after the session
The handoff document
The settings as left, with the reason for each and the date the vendor's terms were read
The connected tools and their scopes, and how to remove one
The workflow's instructions and its test cases, with how to rerun them after a change
The admin tasks of a month: new and leaving staff, the usage export, the seat review
Use by role
| Role | Seat | Seats | Used it | Regular | Median days | Conversations | Workflow runs |
|---|---|---|---|---|---|---|---|
| AP supervisor | Chat | 1 | 1 | 1 | 14 | 44 | 30 |
| AP clerk | Chat | 2 | 0 | 0 | 0 | 0 | 0 |
| Controller | Chat | 1 | 1 | 1 | 11 | 37 | 0 |
| Staff accountant | Chat | 3 | 3 | 3 | 10 | 95 | 0 |
| FP&A analyst | Chat | 2 | 2 | 2 | 13 | 125 | 0 |
| Credit and collections specialist | Chat | 2 | 2 | 2 | 12 | 79 | 42 |
| Payroll specialist | Chat | 1 | 1 | 0 | 6 | 16 | 0 |
| ERP developer | Coding | 2 | 2 | 2 | 18.5 |
Apart from the two AP clerks, one person used the tool without becoming a regular user: the payroll specialist, on 6 working days. Coding assistant seats are counted by days of use; their conversations are not in the chat export.
By week
| Week | Dates | Working days | People | Conversations | Per working day | Workflow runs |
|---|---|---|---|---|---|---|
| 1 | March 4 to March 7 | 4 | 11 | 40 | 10.0 | 9 |
| 2 | March 10 to March 14 | 5 | 11 | 76 | 15.2 | 15 |
| 3 | March 17 to March 21 | 5 | 12 | 104 | 20.8 | 10 |
| 4 | March 24 to March 28 | 5 | 12 | 97 | 19.4 | 25 |
| 5 | March 31 to April 2 | 3 | 12 | 79 | 26.3 | 13 |
Conversations per working day more than doubled from the first week to the last (10.0 to 26.3). The first and last weeks are partial, so the table compares days, not weeks.
Spend
| Chat assistant seats, 12 | $360.00 |
|---|---|
| Coding assistant seats, 2 | $50.00 |
| Coding assistant usage, 61.3 million tokens | $367.53 |
| Total for the period | $777.53 |
| Per person who used it | $64.79 |
| Per regular user | $70.68 |
The two coding assistant seats cost $417.53 with their usage, more than the twelve chat seats together. Usage-billed tools are where a monthly bill moves, so the report shows them line by line; the AI Cost Controls Setup adds limits and alerts where a team needs them.
Prices used in this sample
| Name | Set for this sample | Value |
|---|---|---|
chatSeatMonthly | Chat assistant seat, business plan, per month | $30.00 |
codingSeatMonthly | Coding assistant seat, per month, before usage | $25.00 |
codingUsdPerMtok | Coding assistant usage, blended price per million tokens | $6.00 |
regularDays | Working days of use that make a person a regular user, out of the period's working days | 8 days |
These are not vendor prices. Licenses are billed by the vendor at its price on the day.
What this sample is and is not
- It is a sample, not client work. The team belongs to the synthetic distributor of the Operations and Cost Diagnostic sample, whose AP headcount it shares. No real company or person is described.
- The usage figures are computed by a script from two generated tables: seats (14 rows), usage_log (147 rows), from a fixed seed. On a client's rollout the same report reads the vendor's usage export. Figures generated on 2026-11-21.
- The checklist, the connected tools, the instructions, the test cases, the session and the handoff are the written deliverables as they would be handed over. The findings in the checklist are part of the made-up case.
- No model was run for this sample and no vendor setting was read. Usage counts say how often the tool was used, not whether the work got better; the workflow's test cases are how that is checked.
The same rollout, for your team.
From $4,900, about two weeks. Tell us the team's size, the tool you use or are choosing, and the work it does most.