AI at Tab & Shift: what it can do, and what it costs
Tab & Shift Engineering Ltd. · For the board meeting on June 4, 2026
| Prepared for | Prepared by | Date | Status |
|---|---|---|---|
| The board Chair: Paige Tab | Floor 16 Technologies | May 28, 2026 | For decision |
People at Tab & Shift already use AI: two of the four teams we interviewed paste proposal text into free chat tools. Doing it safely costs less than most boards expect and takes longer than vendors say. We recommend three steps, the boundary first, and ask for four decisions today. None of them is a paid seat for everyone.
What AI can do for us today, and what it can't
AI tools are good at first drafts, summaries and finding things in documents. They are not good at being right without a check. Every example below is Tab & Shift work.
It can
- Draft a proposal from our approved past proposals and the boilerplate library: a first draft in half a day, not two. The bid team still checks and prices it.
- Turn an hour-long Teams meeting into decisions, actions and owners before people leave the room.
- Answer "what is our standard detail for a pile cap?" from our own standards, with a link to the page.
- Read a 60-page contract or specification and list the clauses that matter: liability, payment terms, deadlines.
- Draft a daily field report from a supervisor's notes and photos.
It can't, or shouldn't
- Check, stamp or sign engineering work. An engineer still checks every calculation and drawing, as now.
- Price a bid. It drafts the scope; pricing stays with the Commercial group.
- Be right every time. It can state a wrong figure with confidence, so the person who sends it owns it.
- Tell old from current on its own. With three versions of a policy published, it may quote the oldest.
- Keep a secret we have not locked. It sees what the person asking can open, and today that includes payroll.
The tools, in one line each
| Tool | Good at | What it reads |
|---|---|---|
| Copilot in Microsoft 365 | Email, meetings and documents, inside Word, Outlook, Teams and Excel | The person's own Microsoft 365 content |
| Claude | Long documents, contracts, analysis and writing | SharePoint and OneDrive through a connector, as the person signed in |
| ChatGPT Enterprise | General drafting and analysis, for people already used to ChatGPT | Connected sources, as the person signed in |
The boundary work is the same whichever tool we pick, so the choice can wait for the pilot. Most companies end up with more than one.
What it costs
Year one, in US dollars, rounded. Floor 16's work is one time; licenses repeat every year. Every cost is shown and approved before anything is built or bought.
| Item | Why it is needed | One time | Each year |
|---|---|---|---|
| The boundaryFloor 16: cleanup, owners, labels, the test | Fixes who can open payroll and pricing today, with or without AI | $18K–26K | |
| Business Premiumall 90 staff, up from Business Standard | Labels, encryption and device rules need it | About $10K | |
| Purview add-onabout 15 people in HR, finance and leadership | Labels set by library, and the most sensitive files kept out of Copilot | About $2K | |
| The pilotFloor 16: setup, training, weekly measures | Six weeks on real work, with numbers at the end | $6K–9K | |
| Pilot seats10 paid AI seats | Copilot is bought a year at a time. Claude and ChatGPT cost about the same per seat | About $3K | |
| Rollout trainingFloor 16: a session per team, prompts, office hours | So the seats bought get used | $4K–6K | |
| Rollout seats25 to 30 more people, by role | Only if the pilot clears its bar. Everyone else stays on Copilot Chat | $6K–11K | |
| Total | $28K–41K | $21K–26K |
Seat prices are approximate list prices for a company of 90, about $21 to $30 a person a month, and they change. Final figures come from the reseller and each vendor before anything is bought.
About half of year one is not AI. The boundary and Business Premium fix who can open payroll and pricing today, whether or not AI is ever switched on. If the pilot falls short, that work still stands and the rollout seats are never bought.
What the pilot has to show: 8 of 10 people using it every week by week four, a median of two hours a week saved each, and no boundary incidents. Below that bar, we stop at ten seats.
The risks, and how they are handled
| Risk | What it means | How it is handled |
|---|---|---|
| AI sees too much | AI answers as the person asking. If a field technician can open payroll, so can the AI they use. | The boundary comes first: "Everyone" removed from every site, two owners on each, then a test with the questions we hope it cannot answer, asked as real roles, before anyone gets a seat. |
| Files in the wrong place | A fee proposal copied into a project team is open to the whole team. | Six sensitivity labels, most applied by where a file is saved. HR, payroll, board and pricing labels are encrypted, so tools outside Microsoft cannot read them, and Copilot is set not to use the most sensitive two. |
| Free tools | Company text in free chat tools, on personal accounts, under terms we never agreed. | Company accounts only, with our data kept out of model training by contract. A one-page approved-tools guideline from the day the pilot starts. |
| A wrong answer goes out | A draft with a wrong figure reaches a client. | A person checks and sends everything. Engineering work is checked by an engineer, as it is today. |
| Seats nobody uses | Paid seats sit idle until the renewal. | No seats beyond the pilot until it clears its bar. Seats unused for 30 days move to someone else. |
| The vendor changes | A tool changes what it reads or how it treats labels. | The boundary test is repeated every quarter. The labels work the same whichever tool we use. |
The plan in three steps
Clear out stale and duplicate files, two owners on every site, six labels, then the boundary test. Done when every test question returns only what the person asking should see.
$18K–26K, plus licensesTen people in bids, marketing and project management, on real work. A one-minute form every Friday, and three live bids timed against the last three.
$6K–9K, plus 10 seatsIn waves, a team at a time, with a session on each team's own work. Copilot Chat for all 90, and paid seats for the 35 to 40 roles the pilot shows will use them.
$4K–6K, plus seats by roleBetween steps 2 and 3, the board sees the pilot numbers in November and decides on seats. Nothing beyond the pilot is bought before then.
What we ask the board to decide today
| # | Decision | We recommend | Cost |
|---|---|---|---|
| 1 | Start the boundary work | Approve step 1 now. It is needed for security with or without AI. | $18K–26K once |
| 2 | Move to Business Premium | All 90 staff at the July renewal, plus the Purview add-on for about 15 people. | About $12K a year |
| 3 | Fund the pilot | Grace Lin's bid team, with Priya Shah as sponsor. Released only when the boundary test passes. | $6K–9K once, about $3K a year |
| 4 | Set the rule on free tools | Company content goes only into Tab & Shift's own AI accounts, from the day the pilot starts. | None |
Not asked today: seats for everyone. That comes back to the board in November, with the pilot's numbers.
Three questions to ask any vendor
Every AI vendor will ask for a meeting this year. These three questions separate a serious offer from a sales pitch. Ask for the answers in writing.
What can it see, and who decides?
Does it only answer from what the person asking can already open? Can we keep HR, payroll and board files out completely? Ask to see it on our own files, not a demo.
What happens to our data?
Is anything we send used to train their models? Where is it stored, for how long, and how do we get it deleted? The answer belongs in the contract.
What does it cost after year one, all in?
Seats, the licenses it needs underneath, add-ons and the minimum term. Can we cut seats at renewal, and what happens to the price then?