One team first, then the rest
Tab & Shift Engineering Ltd. · Oct 5, 2026 to Jan 22, 2027
| Sponsor | Prepared by | Date | Status |
|---|---|---|---|
| Priya Shah Director of Operations | Floor 16 Technologies | Sep 28, 2026 | v1.0 · Agreed |
Summary
Sixteen weeks. Ten people use AI on real work for six weeks, and nobody else gets a paid seat until the pilot shows who uses it. Then three waves, and five agents for the jobs that repeat.
This plan starts where the AI Boundary Check ended. Content is current, owned and labeled; the boundary test passed 72 of 72 runs on Sep 24. Everyone at Tab & Shift already has Copilot Chat with their Microsoft 365 license. What this plan decides is who gets more than that, in what order, measured how.
How Tab & Shift got here
The Scan: 41 sites open to everyone, no retention, no labels. Maturity 2.0 of 5.
Retention schedule, 322,000 files archived, 78 sites to 44, two owners each.
"Everyone" removed, six labels, restricted discovery on sensitive sites.
24 prompts, three tools, 72 of 72 passed after six site fixes.
One team, on real work, measured. Then the rest.
The sixteen weeks
Week 1 starts Monday, Oct 5, 2026. Week 16 ends Friday, Jan 22, 2027. Invoice processing continues past week 16.
Who, when, and how we know it worked
| Stage | Weeks | Who | Owner | Success measure |
|---|---|---|---|---|
| Wave 1 · Pilot Bids and Proposals, marketing, four project managers | 1 to 6 | 10 | Grace Lin, Business Development Lead | 8 of 10 use it weekly by week 4. Median 2 hours saved a week. First draft of a proposal in half a day, down from two. No boundary incidents. |
| Gate | 6 | Priya Shah, Marcus Chen | Continue, adjust or stop, on the pilot numbers. Seat plan for waves 2 to 4 agreed. | |
| Wave 2 · Engineering Engineers and drafting | 7 to 9 | 24 | Hannah Brooks, Engineering Manager | Draft reports checked against standards before review. Review comments per report down by a third. |
| Wave 3 · Field Services Supervisors and crews | 10, 11, 14 | 30 | Luis Ortega, Field Services Manager | Supervisors use it for job questions and daily email. Crews use Copilot Chat on their phones; no paid seats. |
| Wave 4 · Office and leadership Finance, HR, admin, leadership | 14 to 16 | 26 | Marcus Chen, CFO | Meeting summaries and month-end commentary drafted. Restricted content stays out of every answer, checked at week 16. |
| Close | 16 | All | Priya Shah | Seats unused for 30 days reassigned. Results written up. Quarterly boundary retest booked. |
How the pilot is measured
A one-minute form every Friday: what you used it for, roughly how long it saved, and anything it showed you that you think you should not have seen. Usage comes from the Microsoft 365 admin reports and Claude's admin console. Three live bids are timed from kickoff to first draft and compared with the last three before the pilot.
Seats, by evidence
Every Tab & Shift user already has Copilot Chat, grounded on the web and the file they have open, at no extra cost. Paid seats add the person's own Microsoft 365 content and the apps. The numbers below are the plan after the gate; nothing beyond the pilot is bought until then.
| Group | People | Copilot Chat | Microsoft 365 Copilot | Claude | Why |
|---|---|---|---|---|---|
| Bids, marketing, project managers | 10 | 10 | 10 | 5 | Drafting from the approved library; long proposals and contracts |
| Engineering | 24 | 24 | 12 | 1 | Report checking; the senior half write most reports |
| Field Services | 30 | 30 | 6 | 0 | Supervisors only; crews use Chat on phones |
| Office and leadership | 26 | 26 | 6 | 0 | Leadership and finance leads |
| Total | 90 | 90 | 34 | 6 | Today: 10 Copilot, 0 Claude |
Seat counts after week 6 are a plan, not a purchase. Pricing is confirmed with Tab & Shift's reseller and with Anthropic at the time.
The ways AI can be deployed
Five ways, and most companies end up using more than one. The choice is made per job, not per vendor.
| Option | Use it when | What it reads | How the boundary applies | Tab & Shift's decision |
|---|---|---|---|---|
| Copilot in Microsoft 365 | The work is in the document, email or meeting in front of you: Word, Outlook, Teams, Excel. | The person's own Microsoft 365 content, at their permissions | Permissions, labels, the Copilot exclusion on Restricted, restricted discovery | Chat for everyone; 34 paid seats by role after the gate |
| Copilot agents | One job repeats against one body of content: policy questions, onboarding, a project's documents. | Only the sites and lists the agent is given, as the person asking | As above, plus the agent's own source list | Policies agent on Tab & Shift Home (week 8); standards agent on Engineering Standards (week 10) |
| Claude Team or Enterprise plan, with the Microsoft 365 connector | Long documents, analysis and writing: proposals, contracts, research, reports. | SharePoint, OneDrive and, if enabled, mail, as the signed-in person | Permissions; encryption keeps Commercial and Restricted content out, tested | 5 seats in the pilot for proposal writers; mail connector off |
| ChatGPT Enterprise | General drafting and analysis with company controls, for teams that already use ChatGPT. | Connected sources, as the signed-in person | Permissions; encryption, tested | Passed the boundary test; held as the alternative if a team prefers it to Claude |
| Azure AI Foundry | A job runs at volume without a person starting it: inbound email, invoices, scanned forms. | Only the sources Floor 16 connects, in Tab & Shift's own Azure subscription | Nothing labeled Commercial or Restricted is connected | Email triage, job sheet reading, invoice processing; pay for what runs |
Free tools stop on Oct 5. The approved-tools guideline goes to all staff with the pilot. Company content goes into Copilot, Claude or ChatGPT Enterprise under Tab & Shift's accounts, and into nothing else. Two of four people interviewed in April said proposal text was going into free chat tools; this is the replacement, not a ban.
Five agents for the jobs that repeat
Each one takes a step nobody wants, and leaves the decision with a person. The person's step is marked in Deep Sky.
1 · Marketing material drafts
Firm · weeks 2 to 6Trigger
A project closes and the client agrees to a case study.The agent
Drafts the case study from the close-out report and site notes, in Tab & Shift's style, from twelve approved examples.The person
Marketing edits; the project manager checks facts; the client approves.Result
Published, labeled Public.2 · Workflow approvals: expense claims
Firm · weeks 8 to 12Trigger
A claim is submitted with receipt photos through the expense form that goes live in October.The agent
Reads the receipts, checks amounts, dates and project codes against the expense policy, and writes a two-line summary with any exceptions.The person
The approver reads the summary and approves, queries or rejects.Result
Approved claims export to payroll weekly.3 · Incoming email triage
After the gate · weeks 9 to 16Trigger
Mail arrives in the info and jobs mailboxes: about 1,400 a month.The agent
Sorts each message (client request, question on a job, supplier, invoice, noise), files it to the project by number, and drafts replies to routine requests.The person
The coordinator reviews the drafts and sends; anything uncertain is left for them.Result
Filed to the project, answered the same day.4 · Process automation: job sheets
After the gate · weeks 11 to 16Trigger
A crew still on paper emails a photo of a handwritten job sheet: about 300 a month.The agent
Reads hours, equipment and project number from the photo and writes a draft row to the register, the same as the phone form does.The person
The supervisor confirms the row, or corrects it.Result
In the weekly billing export; nobody retypes it.5 · Invoice processing
After the gate · from week 14Trigger
A supplier invoice reaches the accounts payable mailbox: about 400 a month.The agent
Reads the invoice, matches the supplier and project, suggests the GL and project codes, and flags duplicates and price differences.The person
The project manager approves; finance posts.Result
Imported to the accounting system, coded.Training and support
| What | Who | When |
|---|---|---|
| A 45-minute session per wave, on that team's own work, with the five prompts that save them the most time | Floor 16 | First day of each wave |
| A prompt library on Tab & Shift Home, by role, added to by the pilot team | Grace Lin, Floor 16 | From week 2 |
| Office hours, 30 minutes a week, for anyone | Floor 16 | Weeks 1 to 16 |
| One champion per team, who answers first and passes on what works | Each wave owner | From each wave start |
| The approved-tools guideline: which tools, what you own in the output, how to report something you should not have seen | Priya Shah, Elena Rossi | Oct 5 |
Risks and how they are handled
| Risk | What would happen | Control |
|---|---|---|
| The boundary drifts | Someone saves a pricing or HR file in the wrong place and AI finds it. | Content rules relabel strays; retest before wave 2 and every quarter; "report it" in the guideline |
| Seats bought before use | Paid seats sit idle until the renewal. | Nothing beyond the pilot bought before the gate; 30-day unused seats reassigned |
| People trust a wrong answer | A draft with a wrong figure goes to a client. | Every agent leaves the decision with a person; the guideline says the sender owns the output |
| The pilot is too easy | Results from enthusiasts do not hold for wave 3. | The pilot includes four project managers who did not ask for it |
| Vendor changes | A tool changes how it handles labeled files. | Quarterly boundary retest; the label design does not depend on any one vendor |